Study design and study population
The Prevention and Incidence of Asthma and Mite Allergy (PIAMA) study is a birth cohort study in the Netherlands that started in March 1996. During their pregnancy 7 862 mothers were recruited from the general population in three different regions of the Netherlands. The recruitment of the mothers started during their first visit to a prenatal health clinic with the help of 52 midwifes. No exclusion criteria were used. In total 4 146 mothers agreed to participate and gave written informed consent. All children were born in 1996 or 1997. There were 183 mothers lost to follow-up during the first data collection on the child. For this reason the baseline study population consisted of 3 963 children. The first questionnaire was completed by the parents during pregnancy. The second questionnaire was completed by the parents when the child was 3 months old, and subsequently, every year until the child was 8 years old. In 2008, at the child’s age of 11 years, the parents received a questionnaire again. The study protocol was approved by the Medical Ethics Committees of the participating institutes (Rotterdam, MEC 132.636/1994/39 and 137.326/1994/130; Groningen, MEC 94/08/92; Utrecht, MEC-TNO 95/50). Detailed descriptions of the study design has been published previously [4, 5]. For our analyses, children were excluded if data were missing on all BMI measurements (n = 32), bedroom and living room temperature (n = 89), or confounders (n = 497). Furthermore, children with reported values for living room temperature lower than 10°C were also excluded (n = 2). After exclusion, data from 3 343 children were available for the analyses of the original data.
BMI
Child’s height (in cm) and weight (in kg) were reported by the parents at the age of 3 months, 1 year, 2 years, 3 years, 4 years, 5 years, 6 years, 7 years, 8 years and 11 years. Parents were asked to report the child’s weight and height measured by a medical professional, if this measurement was within the last 3 months. Otherwise parents were asked to measure weight and height of the child themselves without shoes and heavy clothes. Based on data on child’s height and weight the continuous outcome variable BMI z-score was calculated. First, BMI was calculated as weight divided by height squared in meters (kg/m2). BMI was not calculated if data on weight or height were missing and/or if the period between the measurements of weight and height was longer than 60 days. Subsequently, gender and age-specific BMI z-scores were calculated by using the reference growth curves of the Dutch fourth nationwide growth study that was performed in 1997 [6].
Indoor temperature
Winter indoor temperature in both living room and bedroom was reported by the parents in the first year of life. The following open question on temperature was asked for living room and bedroom: “At what temperature do you keep your room in winter?” If people did not use the heating, they could choose the additional answer “heating is not used”. Both living room and bedroom temperature were divided into two groups: “heating used” versus “heating not used”. Temperatures within the category “heating used” were used as a continuous variable.
In a subgroup of children (n = 104) bedroom temperature was measured with data loggers every 30 minutes during a period of approximately 2 weeks. These measurements took place in November, December, January and February of the years 1997, 1998 and 1999 and were used to calculate a mean measured bedroom temperature for each child in this subgroup.
Statistical analyses
Data were analyzed using SAS 9.2 (SAS Institute, Inc., Cary, NC, USA). Descriptive statistics were used to describe the characteristics of the study population. The regression analyses described in the following sections were performed with generalized estimating equations (GEE) models with a nine-dependent correlation matrix. These models were used to take into account correlations between the repeated measurements of BMI within the same subject.
Both the overall association and age-specific associations between reported living room temperature and BMI were analyzed for children who lived in houses where the heating was used in the living room. These analyses were repeated for children who did not move during the entire 11 years study period, the non-movers. The latter was done because indoor temperature was reported at baseline and moving could contribute to a change in heating habits.
Like for living room temperature, both the overall association and age-specific associations between reported bedroom temperature and BMI were analyzed for children who used the heating in their bedroom. These analyses were also repeated for children who did not move during the study, the non-movers. In addition, the overall association and age-specific associations between reported bedroom temperature as dichotomous variable (“heating used” versus “heating not used”) and BMI were also analyzed.
For the subgroup of children with objectively measured bedroom temperatures available (with data loggers), the overall association between measured bedroom temperature and BMI was analyzed. In addition, for children in the subgroup with both a reported and a measured bedroom temperature available, the Pearson correlation coefficient was calculated between these temperatures.
GEE models were stratified by gender because the development of the BMI in childhood is different between boys and girls. Furthermore, GEE models were adjusted for the confounders birth weight, maternal education level (low, intermediate, high), maternal overweight before pregnancy (BMI < 25 kg/m2 and BMI ≥ 25 kg/m2), indoor smoking (yes/no) and breastfeeding (yes/no). Indoor smoking and breastfeeding were used as lifestyle indicators. Consideration of potential confounders was based on a difference greater than 10% in effect estimate between crude and adjusted model and/or theoretical considerations. Maternal overweight before pregnancy could be a confounder in the association between indoor temperature and BMI, but could also be an effect of indoor temperature. The latter means that maternal overweight before pregnancy would not be a determinant of indoor temperature and could not be a confounder in this case. For this reason two different adjusted models were analyzed; one model with adjustment for maternal overweight before pregnancy and one model without this adjustment.
Multiple imputations
In the original dataset (n = 3 343), the percentage of missing values on the child’s BMI ranged from 6.4% at the age of 1 year to 36.4% at the age of 11 years. Since missing data from general population studies are unlikely to be missing completely at random (MCAR), our analyses might lead to biased results because of selection bias. For this reason, missing data on all variables were multiply imputed by using the Multivariate Imputation by Chained Equations (MICE) procedure [7] in the statistical program R 2.13.1 [8]. Results of the analysis on each of the 20 imputed datasets, of 3 963 children each, were combined using PROC MIANALYZE in SAS 9.2.
Analyses were first done with the original dataset consisting of children with complete data on living room and/or bedroom and confounders (n = 3 343). Analyses were repeated with 20 imputed datasets of 3 963 children each. In this article, results from the analyses of the imputed datasets were reported and compared with results from the analyses of the original dataset with complete data on living room and/or bedroom and confounders.