Study selection

To identify datasets with repeated measures of child TSE that included both parental reports and objective measures, we searched the literature for original trials of interventions that attempted to reduce child exposure to tobacco smoke, involved participants who were the parent(s) (mother, father or both parents) of children aged 0–18 years; and in which the study outcome was the child’s TSE as reported by a parent and as measured using one or more biomarkers. We searched Web of Science, MEDLINE, BIOSIS Previews, Journal Citation Reports, and the Cochrane library, for articles published in English through the end of 2012. Search terms used with all databases are presented in Supplementary file S1.

Of 61 studies identified, most were longitudinal, with outcomes measured at several time points; some included replicate objective measurements at the same time point. We included studies with at least 3 longitudinal time points, at least 2 parental reports of their smoking behavior and/or child’s exposure at each time point, and sample sizes larger than 150. We were able to obtain raw data from 3 studies: Hovell et al.¹⁸ (2009), Kalkbrenner et al.¹⁹ (2010), and Wilson et al.²⁰ (2011) (ClinicalTrials. gov Identifier: NCT00217958). We later included an additional study conducted by authors of this article, Rosen et al.²¹ (2021) (ClinicalTrials.gov Identifier: NCT02867241), although it assessed TSE at only two time points. These four studies will be referred to as the studies of Hovell, Kalkbrenner, Wilson, and Rosen, for brevity.

We explored the relationships between parent-reported and objective measures of child TSE from all children in the studies. We were particularly interested in understanding which, if any, of the parent-reported measures were most highly correlated with the objective measures. We were also interested in differences in correlations between measures, over time and by intervention arm. We assessed the evidence for a ‘learning curve’ whereby parents might report exposure more ‘accurately’ (that is, more consistently with the objective measures) over time, as a result of having been sensitized to the questions in the early trial period. Finally, in the two trials that used replicate objective measures, we examined within- and between-subject variation to understand whether replication yields information that provides a more useful exposure assessment than a single measure alone.

All the objective measures were analyzed after log transformations, as is common in the literature, due to their skewed distributions. We used data as provided to us by the authors of the studies.

Statistical analyses

We performed the following analyses: associations between objective measures and parental report variables summarized by Pearson correlation coefficients. We used linear regression to discover which parent-reported measures best predicted the objective measure values obtained at the same time points. The regression models included time, group and their interaction and other potential explanatory variables: child-related, parent-related, sociodemographic and environmental variables (Supplementary file Tables S3–S10). We used forward selection to determine which parent-reported measures to include in the model and 10-fold cross-validation to determine which set of parent-reported measures best predicted the objective measure²².

We were interested in understanding if there was a learning curve over time, by which study participation led to increased parental awareness of TSE, and hence to parental responses more closely correlated with the objective measures. To assess trends in correlations over time, we regressed Fisher transformed²³ correlation coefficients, z′=12 ln (1+r1−r) against time.

To assess the value of replicate objective measures, we used restricted maximum likelihood to estimate within- and between-subject variation^18,20.

In the analyses that included all time points, the longitudinal measures for each subject cause dependencies that render standard p-values inaccurate. For those analyses, we used bootstrap confidence intervals (CIs) for the correlations, with bootstrap samples taken at the subject level²⁴.

The Wilson and the Hovell data included three baseline measurements of urinary cotinine on samples obtained at intervals of approximately 1–2 weeks (Wilson) and 1 week (Hovell). We used two analytic approaches, first computing the average of the three measurements followed by a log transformation, and secondly using each of the three baseline measurements independently. In subsequent analyses, the log of the averaged triplicate was used.

All studies asked multiple questions of the parent(s) about their and/or others’ smoking (overall, in the home, or in the car) and the child’s exposure to that smoking, but the sets of questions differed among the studies, and the specific wording differed in potentially important ways.

We use abbreviations for variable names (see the Abbreviations list above). For example, NCIGS (number of cigarettes) denotes heaviness of smoking among residents of the household. There were differences in the actual definition in different studies. The questions and definitions for each study are presented in Supplementary file Table S2. LSCot was defined as log serum cotinine, LUCot as log urinary cotinine, LHCot as log hair cotinine, LHNic as log hair nicotine, and LANic as log air nicotine.

Association between objective measures and parental reports

Overview

The correlation coefficients between objective measures and parent-reported measures ranged from moderate (highest: r=0.41) to essentially zero. Among the parent-reported measures, SMOKERULE had the highest correlations with the objective measure (r=0.41) followed by NCIGSHOME (0.39) or NCIGS (r=0.37). Objective measures LSCot and LUCot had better correlations with all the parent-reported measures than did the objective measures derived from hair (LSCot: range 0.26–0.37; LUCot: range 0.12–0.41; LHNic: range 0.04–0.21; LHCot: range 0.05–0.15) (Figures 1 and 2, and Table 2).

Figure 1

Correlations between objective measures and parent-reported measures across both intervention and control groups at all assessment points for each study. (a) Correlations between objective and parent-reported measures; (b) Correlations between objective environmental measure (LANic) and parent report variables, by study

Figure 2

Comparing the correlations between objective measures and parent-reported measures across both intervention and control groups at all assessment points by parental measure of exposure

Effect of using a single baseline reading of LUCot versus the average of 3 readings

Wilson and Hovell took multiple baseline measurements of urinary cotinine. In Wilson’s study similar correlations between urinary cotinine and parent-reported measures were observed when the average of the baseline cotinine measures was used as when the individual cotinine measures were used. In Hovell’s study, we found slightly, but not significantly, larger correlations between the average of the baseline cotinine measures and parent-reported measures than between the individual cotinine measures and the parent-reported measures. For example, the correlations of the 3 repeats of LUCot at baseline with SMOKERULE for the 3 separate measurements at baseline were r=0.42, 0.46, and 0.44, respectively. By contrast, after averaging the three measurements, the correlation was r=0.48 (Supplementary file Table S2).

Prediction of objective measures by parental report

Table 3 summarizes the cross-validation results according to study and objective measure. The first line for each objective measure gives the cross-validation mean square when ignoring all the parental report variables. Subsequent lines give the mean square when one or more parental report is used to predict the objective measure result. Comparison with the top line quantifies what fraction of the crossvalidated variation was explained by the regression model (Supplementary file Tables S3–S10).

When using LSCot, NCIGSHOME best predicted the objective measure (R²=0.139). When using LUCot, SMOKERULE and NCIGSHOME or NCIGS gave the best predictions (R²=0.165, 0.134, 0.131, respectively). LHCot and LHNic had weak correlations with the parental reports (max R²=0.021 and 0.043, respectively) and were not accurately predicted by any combination of them. The accuracy in prediction of objective measures from parental reports was stable across time and group, as evidenced by their similar mean square errors (MSE).

The strongest predictive relationship was found for urinary cotinine in Hovell’s study and included the variables home smoking rules and total number of cigarettes smoked in the home (R²=0.209). Additional explanatory variables found to be significantly related to the objective measures in more than one study were: gender (female gender was associated with greater exposure), number of rooms in the house (more rooms were associated with lower exposure), assessment time point (later assessment points were associated with lower levels), race (the objective measure was higher in Black than in White participants) and income (negative association).

Analysis of change in correlations over time

We did not find any consistent time trend in correlations between parental report variables and objective measures in the different studies, and time was generally not a statistically significant predictor. For details see Supplementary file Table S2.

Assessing the value of replicate measures

Supplementary file Table S11 shows the estimated within- and between-subject variances in the Hovell and Wilson studies for multiple baseline measurements taken within a reasonably short span of time with no intervening intervention (e.g. three measurements of LUCot at baseline in both Hovell and Wilson). The between-subject variances were similar (0.65 and 0.57, respectively). However, the within-subject variances differed substantially (0.22 and 3.15, respectively).

Regarding correlations between objective measures and parental report

Poor synchronization between the time period measured by a given biomarker, and the time period assessed by parental reports, could partially explain low correlations. Our observations regarding this issue differed across studies. For example, Kalkbrenner used serum cotinine (24–48 hours exposure) and hair cotinine (several months exposure). Their questionnaire referred to the child’s exposure in the past 24 hours, suggesting, a priori, that there should be a higher correlation with serum cotinine than with hair cotinine. Indeed, this was borne out empirically (r=0.37, p<0.001 with LSCot; and r=0.08, p=0.066 with LHCot). By contrast, the study of Rosen found that hair nicotine had higher correlations with exposure during the previous day (r=0.21) than in the recent month (0.04–0.06).

Biomarker expression may differ across individuals. For example, hair nicotine has been found to be associated with hair color²⁵.

It may be that parental reporting of the number of cigarettes smoked is unreliable and/or systematically biased toward over-/under-reporting. However, even if parents are reporting reliably, the questions may not be sufficiently detailed to capture the full range of possible situations and so assess true exposure. For example, assuming that a reasonably strong association exists between the number of cigarettes smoked in a home and child exposure, the observed correlation may be reduced because individual children spend varying amounts of time indoors – and perhaps less so as they mature. Home architecture may differ, ventilation may differ, and parents may smoke in different areas or in different amounts. This complexity demonstrates the importance of valid biomarkers for exposure measurement and the importance of developing questions which better capture exposure.

Similar modest correlations were found in the literature. Tang et al.¹², in their summary of the literature, reported correlations of 0.22–0.69 between hair nicotine and parental report (see also Supplementary file Table S5), -0.02–0.77 for urinary cotinine (see also Supplementary file Table S6), and 0.34–0.35 for serum cotinine (see also Supplementary file Table S7). Modest correlation was found between number of cigarettes smoked per day and air nicotine r=0.36, and r=0.44 for urinary cotinine²⁶, and lower correlation between number of cigarettes smoked by the father and hair nicotine from the father, their spouse and the child (r=0.13; r=0.07, and r=0.03, respectively)¹⁵. Similar results were found in infants, with correlations from r=0.35 with plasma nicotine and number of days the infant was exposed to TSE, to a maximum of r=0.66 between hair nicotine and number of cigarettes smoked per day²⁷. Another recent study found a large range of correlations, from -0.2 to +0.4, which were not statistically significant, between parent-reported measures and air and hair nicotine²⁸. However, one study found higher and statistically significant correlations between maternal hair nicotine and parent-reported measures (r=0.46–0.74)²⁹.

Other fields also suffer from low correlations between objective measures and reported exposure. For instance, in nutritional epidemiology, most correlations that Trabulsi and Schoeller³⁰ found between dietary energy intake and doubly labeled water were not statistically significant and ranged from r=0.06 to r=0.86. Other studies also reported low correlations: Freedman et al.³¹ found that the correlations between reported (via three 24-hour recall food frequency questionnaires) and measured energy intake ranged from r=0.21 to r=0.49. Among females, the correlations between data from food recall questionnaires and measured sodium:potassium ratio ranged from r=0.15 to r=0.64³².

Our analysis found that urinary and serum cotinine had higher correlations with parental report than did hair nicotine. Our results differ from those of Al-Delaimy et al.³³, who found parental report to be more highly correlated with hair nicotine than with urinary cotinine, leading them to conclude that hair nicotine was a ‘more precise biomarker of exposure than urinary cotinine’.

Learning curve

Prior to performing the analyses, we had hypothesized that there might be a ‘learning curve’ in which parents would report child exposure more accurately over the course of a study. In principle, more accurate reporting should lead to higher correlations at later times. No evidence of increased correlations between reported and observed measures over the course of the study was found in any of the four studies when the data from all groups were examined together. When examined separately, the results were inconsistent. Our conclusion does not support the hypothesis that correlations increase over time.

Regarding replicate samples of biomarkers

Exposure of a child may vary significantly from one day to the next. Replicates across a short time window enable us to assess the variation and to provide a more precise assessment of typical exposure³⁴. As with any measure, results will also be affected by other sources of measurement variation such as sampling techniques and laboratory precision, including limits of detection and quantification^35,36.

Many factors need to be considered in deciding whether to take short-term replicates. Recruitment, screening, and assessment of additional subjects is costly, especially in trials that examine individually delivered behavioral interventions. These must be weighed against the cost for collection and analysis of replicate samples for each subject. Further, relative costs may vary between countries and research settings.

An important consideration in intervention trials is to maximize power for finding effects on outcome. Greater power may be achieved by maximizing sample size, while taking a single objective measure from each participant at baseline and at the end of the study, or by reducing within-subject variance in the exposure outcome measure. Careful evaluation is needed to assess the trade-off between the expense of the replications and the expense of recruitment and trial participation, to achieve the desired power for study comparisons. Bitan et al.³⁷ provided a design framework for studies with both objective measures and self-reports and found that excessive replication typically compromises power.

We think that decisions on replication should relate to the relative sizes of the inter-participant and intraparticipant variation. It is noteworthy that we found a large difference in intra-participant variation between the Hovell and Wilson studies, although both assessed the same biomarker. This could reflect differences between study populations. Hovell included healthy children from poor families in San Diego California whose mothers reported smoking ≥10 cigarettes per week, whereas Wilson included asthmatic children from the San Francisco Bay area who had minimum urinary cotinine of 10 ng/mL and no socioeconomic targeting. The large variance in Wilson’s study makes replicates more attractive, but this differs from Hovell’s study, in which the intra-subject variation was considerably smaller. Good estimates of these components are required at the study design stage.

We disagree with the recommendation of Matt et al.³⁸ that one should take enough replicate samples so that the within-subject averages have a variance that is roughly 1% of the total variance. This reasoning led them to recommend as many as 9–13 replicates of urinary cotinine. We are convinced that this criterion sets much too high a bar for measurement precision at the individual level and, if followed, would lead to wasted resources and an unjustified burden on research subjects.

Strengths and limitations

A major strength of this research was its analysis of raw data from multiple studies. Our research included four non-randomly chosen intervention trials. Each of these trials is one example of how researchers currently approach the issue of child TSE assessment in field conditions (as opposed to in a laboratory setting, where a validation study might be conducted). Although the studies were all published prior to 2012, they include all the primary biomarkers currently in use for measuring passive exposure to tobacco smoke. A limitation is the small number of included studies and the fact that the questions presented to parents were not consistently asked in every study, and even reasonably parallel questions had variations in wording that would plausibly affect parental reporting. This may well be reflective of the variation in research practices in this field, given the lack of standardization of questions. Also, some variables that were available in the various datasets, but not presented in the previous publications, were not analyzed. It is possible that since these studies were conducted, there have been improvements in analytical methods, reducing lower limits of detection and providing more accurate biomarker measurement. If so, recent studies of biomarkers may provide more accurate measurements of exposure. These four studies cannot be considered to be representative of all research in the field, and the fact that no single trial included all possible objective measures and questions means we are unable to make definitive statements about advantages of different querying methods.

Implications for further research

Biomarkers

Parental reports

Correlations

The low correlations between biomarkers and parental reports may imply that these measures reflect different aspects of exposure. Methods have recently been developed to produce improved measures of exposure that combine the information from objective measurements and self-reports⁴⁰. This is a promising direction which needs to be investigated further. Another benefit of further research is to assess the impact of improvement in analytical methods.

Further meta-analysis of more recent studies that involved both objective measurement of exposure and parent-reported exposure also is needed, focusing particularly on studies that used comparable questions to elicit parent report. Such eta-analysis might also include observational studies (cross-sectional or longitudinal), as well as intervention trials.