Investigating indicators and determinants of asthma in young adults

Abstract

Background:

In epidemiological studies on asthma determinants an extreme variability in results exists, probably due to different criteria utilised for defining of an asthma ‘case’ and for measuring determinants. We aimed to assess multiple indicators and multiple determinants of asthma in young adults by applying latent variable mixture models (LVMMs), a novel statistical modelling with hidden (or latent) variables.

Methods:

We consider the pooled data of 1103 subjects (aged 20–44 years) from the three Italian centres of the European Community Respiratory Health Survey (ECRHS 1), a standardised database. Underlying multiple asthma indicators (clinicians’ diagnosis, self-report symptoms, respiratory trials) both a latent two-class of asthma syndrome, and three continuous latent variables (severity of diagnosed asthma, severity of asthma symptoms, and severity of respiratory function) were investigated.

Results:

Family history was the more relevant predictor of the two-class of asthma syndrome with a risk increase of about 60% per 1 relative with early life events (OR = 1.60, 95% CI: 1.30–1.97). Smoking, active and passive, are predictive for the indicators of severity of asthma symptoms. On average the risk increase of about 10% (OR = 1.10, 95%CI: 1.01–1.20) either per 1 source point of environmental tobacco smoke (ETS) or per 1 packet a day per 10 years. While, the risk of the indicators of both severity of asthma symptoms (OR = 1.59, 95%CI: 1.23–2.06) and severity of respiratory function (OR = 1.37, 95%CI: 1.03–1.82) increase in women compared to men, the risk of the indicators of severity of diagnosed asthma (OR = 0.57, 95%CI: 0.35–0.91) decreases.

Conclusions:

Considering latent modelling perspective for formulating plausible hypotheses in asthma research, this study highlighted that the host (genetic) component measured as number of relatives with life-events of asthma and/or allergies seems to be the primary determinants of overall observed asthma indicators summarised by hidden two-class of asthma syndrome. Furthermore, a secondary (or trigger) role of smoking on the continuous latent variable of severity of asthma symptoms, and a gender reversal effect were suggested.

Appendix: Interpreting Parameters in LVMM

The restricted LVMM of the paths (1)–(4) represented by arrows in Figure 1 is a special model within the general latent variable modelling framework [cf. for example 50] involving both categorical and continuous latent variables.

Let c denote the latent categorical variable designed to capture clusters within an overall population with 2 classes (0/1). Let $f_1 ,f_2 , \ldots ,f_R$ denote the continuous latent variables representing the theoretical constructs underlying the observed binary (0/1) indicators, $u_1 ,u_2 , \ldots ,u_I$. Let d denote the binary (0/1) clinician’s diagnosis. Finally, let $x_1 ,x_2 , \ldots ,x_J$ and $z_1 ,z_2 , \ldots ,z_K$ denote the observed categorical or continuous determinants and confounding variables, respectively.

The model consists of different logistic/linear regressions, and in compact form can be written using four equations: u related to f and z; f related to c and x; and c related to x, and finally the model incorporates the logistic regression thresholds of the clinicians’ diagnosis on the latent class membership, i.e. for i = 1,...,I; r = 1,...,R and c = 0,1:

$$ \eqalign{ \log \frac{{\Pr (u_i = 1|f,z)}} {{\Pr (u_i = 0|f,z)}} = \alpha _i + \lambda _{ir} f_r \cr \hskip7.2pc\quad+ \beta _{i1} z_1 + \beta _{i2} z_2 + \cdots + \beta _{iK} z_K \cr \cr f_r = \alpha _{rc} + \beta _{r1} x_1 + \beta _{r2} x_2 + \cdots + \beta _{rJ} x_J + e_r \cr \log \frac{{\Pr (c = 1|x)}} {{\Pr (c = 0|x)}} = \alpha _c + \beta _1 x_1 + \beta _2 x_2 + \cdots + \beta _J x_J \cr \log \frac{{\Pr (d = 1|c)}} {{\Pr (d = 0|c)}} = \alpha _{dc} \cr} $$

The intercept/mean/threshold terms, α are latent class-specific effects, where the first class is a reference class with coefficients equal to zero, $\alpha _{j0} = 0,\;<$> <$>\;\alpha _0 = 0,\;\;{\text{and }}\,\alpha _{d0} = 0$.

Considering the interpretation of regression coefficients in the path relationships of Figure 1, the latent two-class-clinicians’ diagnosis path (2) is represented by the logistic regression thresholds ( $\alpha _{dc}$) of V3 = clinicians’ diagnosis on the latent class membership (C = asthma syndrome); these thresholds represent sensitivity (specificity) and false positive (false negative) of the clinicians’ diagnosis. The latent class-continuous factor path (2) is represented by the varying across-class factor means ( $\alpha _{rc}$) of intermediate continuous latent variables (F1 = severity of diagnosed asthma, F2 = severity of asthma symptoms, F3 = severity of respiratory function) underlying the binary observed indicators (V6–V19 of Table 1). Assuming the factor loadings of path (1) equal to one $(\lambda _{ir} = 1)$, these factor means can also be interpreted as the log-odds ratios of the positive (yes = 1) responses to the observed indicators set of that factor, and the antilogs of these factor means represents the odds ratios of path (2)→(1).

The determinant–latent class path (3) is represented by the logistic regression coefficients ( $\beta _j$) of latent two-class (C = asthma syndrome) on the observed determinants (V4 = sex, V20 = BMI, V21 = Family history, V22 = Active smoking, V23 = Passive smoking V24 = Respiratory infections, V25 = Having had cats/dogs, V26 = Early exposure to older children, V27 = Occupational exposure, and V28 = Mould in the house), these regression coefficients represent the log-odds of the asthmatic class as compared to the non-asthmatic class for a unit increase in the observed determinants, and the antilogs represent the odds ratios. The determinant-factor path (4) is represented by the linear regression coefficients ( $\beta _{rj}$) of intermediate continuous latent variables (F1 = severity of diagnosed asthma, F2 = severity of asthma symptoms, F3 = severity of impaired respiratory function) on the observed determinants. Assuming the factor loadings of path (1) to be equal to one $(\lambda _{ir} = 1)$, these regression coefficients can also be interpreted as the log-odds ratios of the positive (yes = 1) responses to the observed indicators set of that factor, and the antilogs of these regression coefficients represent the odds ratios of path (4)→(1). Finally, the confounding variables-indicator path (5) is represented by the logistic regression coefficients ( $\beta _{ik}$) of the observed binary indicators (V6–V19) on the observed confounding variables (V2 = site and V4 = age), so the odds ratios of the previous paths are also adjusted by confounding variables.