Study design.

Data for the present study were abstracted from an ongoing study of the influence of indoor environmental factors on respiratory illness during the first 2 years of life. The study began in 1997 in the province of Prince Edward Island, Canada, which has a population of approximately 150,000. The study was approved by the ethics review boards of the Ottawa Hospital and the Health Protection Branch of the Canadian government. Recruitment occurred during the late autumn and winter (cold season) of each year when the ground was frozen. Because of resource constraints, we recruited approximately 60 consecutive newborns each year. All physicians who practice obstetrics in the province participated in recruitment. Women in the third trimester of pregnancy received letters from the physicians’ offices describing the study and requesting participation. Interested women were contacted by telephone to obtain informed consent. Excluded from the study were babies born > 4 weeks premature, those with neonatal respiratory difficulties requiring prolonged hospitalization at birth, and those whose families expected to change residence within 2 years of birth. Only one child per household was studied. Baseline information was obtained on sociodemographics and family history. The participating parents maintained a daily symptom diary from birth until 2 years or until the study ended, on large multipurpose calendars. Each study family was phoned twice monthly to document information from the diary. If parents had omitted recording symptoms on a daily basis, they provided information for the previous 2 weeks based on recall. Parental reporting of child care attendance was also recorded every 2 weeks.

Definition of respiratory illness.

We adapted the method of Samet et al. (1992) to define a respiratory illness episode, the purpose being to identify discrete acute illnesses as opposed to persistent ongoing symptoms, such as a chronically runny nose. We defined the beginning of an illness episode as 2 consecutive days with any one of the four following symptoms: stuffy nose, cough, wheeze, and shortness of breath. The illness episode starts on the first of these 2 consecutive days and ends when there are 2 consecutive days with none of these symptoms, the last day of the illness episode being the last day with a symptom.

The primary outcome of interest was the number of illness episodes prorated on an annual basis (number multiplied by 365/days of observation). A secondary outcome measure, illness days, was defined as the sum of all days occurring within illness episodes, also prorated on an annual basis. For example, if a child had two illness episodes each lasting 3 days, six illness days would be assigned. If a child had two illness episodes each lasting 5 days, 10 illness days would be assigned.

Air sampling for endotoxin and ergosterol.

Sampling was done within the first year of birth, and for 81%, within the first 4 months. Air from the child’s bedroom was sampled for both endotoxin and ergosterol through a three-piece cartridge equipped with a polycarbonate filter for approximately 5–7 days with a Buck model SS sampling pump (AP Buck, Orlando, FL, USA) calibrated at 2 L/min at the beginning and end of the sampling. Forty-eight hours after the pump was installed, the flow rate was checked to ensure it was within 5% of the initial reading. Very high dust concentrations can clog the filter and reduce the pump flow. If this happened, the pump was stopped and air endotoxin concentration was calculated based on the reduced sampling time. The total volume of air sampled ranged from 6.0 to 23.9 m³, limited by the need to maintain an acceptable flow rate. Once collected, the cartridges and filters were sealed in new plastic bags and stored at room temperature under dry conditions.

For endotoxin analysis, the filters were extracted with depyrogenated water (LRW, Associates of Cape Cod Ltd., East Falmouth, MA, USA) assisted by sonication. Samples taken before April 2001 were analyzed by the Limulus amoebocyte lysate (LAL) assay gel clot method from Associates of Cape Cod. The detection limit was 0.25 EU/filter. All subsequent analyses were done by the LAL chromogenic method, also from Associates of Cape Cod. The detection limit was 0.1 EU/filter. All analyses were performed by the same analyst in the same laboratory according to the manufacturer’s instructions. Apart from differences in the lower limit of detection, the two methods gave similar results.

We analyzed ergosterol, an indicator of fungal biomass, by gas chromatography and mass spectroscopy as described by Foto et al. (2005). The volume of air sampled ranged from 10.7 to 25.4 m³. Ergosterol was determined using an Agilent model 5973 quadrupole mass spectrometer (Agilent Technologies, Inc., Palo Alto, CA, USA) operating in the electron ionization mode at 70 eV. Compounds were separated on an Agilent 6890 series gas chromatograph equipped with a 30 m × 0.25 mm inner diameter × 0.25 μm ZB-5 capillary column. The concentration of ergosterol was determined against an authentic external standard in the selective ion mode using m/z 363 and 396. The detection limit in selective ion mode was 4.5 ± 0.6 ng/mL. Ergosterol standard (Sigma Chemical Company, St. Louis, MO, USA) was recrystalized, freeze-dried, and stored at 4°C.

Definition of covariates other than ergosterol.

During the postnatal interviews, several characteristics were recorded every 2 weeks for a period of 2 years: the presence of furry or feathered pets in the house, the presence of smokers inside the house, whether the baby was breast-fed, and the number of hours per week the child was cared for outside the home. Based on this information, we created a summary variable for each of the characteristics. For the first three characteristics, we calculated the percentage of postnatal interviews where the characteristic was declared. For example, if 20 of 50 postnatal interviews mentioned the presence of pets inside the house for a particular child, the value of the pet variable would be 0.4. The last characteristic, the number of hours per week that the baby was cared for outside the home, was averaged over the entire 2-year period to create the child-care variable. For presentation, we then categorized some summary variables. The pet variable was categorized into "never declared pets," "sometimes declared pets," and "always declared pets," which divided responses almost equally into thirds. The exposure-to-smoke variable was categorized into terciles. The breast-feeding and the child-care variables were kept as continuous data. The age variable was defined as the age of the child in the middle of the span of the follow-up period. "Mold surface area" refers to the total surface area of the home with mold visible to trained home inspectors.

Statistical analysis.

We tested the association between the number of illness episodes per year and airborne bedroom endotoxin concentration using multiple linear regression analysis. Valid endotoxin results were obtained for 351 houses. A total of 19 homes were excluded from the analysis—15 because of missing temperature data and four because of missing income data—leaving 332 homes for analysis. Eleven babies of 332 (3.3%) exited the study before turning 2 years of age (mean age, 1.1 years), and 56 babies had not reached 2 years of age (mean 1.7 years) by the last day of data collection used for this analysis. This is an ongoing study with children entering and exiting at different times. The illness episodes and covariates were annualized and thus adjusted for duration of follow-up. To test the effect of the 11 babies who exited early, we repeated the analysis with and without them, and no differences were found in the illness–endotoxin association.

Mold surface area was expressed as ranks from highest to lowest. Endotoxin values followed a log-normal distribution, so they were log-transformed. A multiple linear regression model with the number of respiratory illness episodes per year as the dependent variable and the natural logarithm of endotoxin concentration as the primary independent variable of interest was built with the stepwise method. A categorical variable—the year of sample collection—was added to the model to account for any seasonal variations in illness from year to year and also the change in the lower limit of detection of the endotoxin analytic technique after 2001. Endotoxin and year of home sampling were held in the model along with any variables with p-values < 0.10, resulting in the final model, with variables endotoxin, year of home sampling, temperature, age, mean hour per week that the baby was cared for more than 1 day a week outside the home, percentage of postnatal interviews in which the baby was breast-fed, income, and categorized percentage of postnatal interviews where smokers were declared in home.

A potential confounder was defined as a variable (Tables 1 and ​and2)2) that, if added to the model, would change the parameter (β) of the natural logarithm of endotoxin by > 10%. No confounders were found for the illness episodes model. The residuals from the regression equation were normally distributed (Shapiro-Wilk statistic = 0.9932, p = 0.1362). We also examined the homogeneity of variance assumption, and the chart of residuals against predicted values showed no particular pattern. Interactions biologically plausible were also tested, and none were found to be statistically significant at the 5% level. We found no evidence of interaction between allergies or asthma in parents and endotoxin. Careful examination of each of the partial residual plots (i.e., the component-plus-residual plot) did not reveal any sign of nonlinearity in the relationship between illnesses and air endotoxin. Further, adding a square term for endotoxin did not improve significantly the R² of the model.

The endotoxin measurement was made only at the beginning of the 2-year follow-up. To determine the robustness of the endotoxin–illness association, we measured it at several time points between the initial endotoxin measurement and symptom assessment. We would assume that a true causal association would remain stable or weaken over time. If the association increased or fluctuated randomly with time of follow-up, this would reduce the probability of a causal association. We measured the results from 90-, 180-, 270-, 360-, 450-, 540-, 630-, and 720-day windows around the time of endotoxin sampling. The regression model obtained previously for a 2-year period was applied to each of these windows. The β-coefficient for the effect of the natural logarithm of endotoxin on illness episodes and total illness days along with its 95% confidence interval were graphed against the size of the window.