No clear link between indoor fuel use and eczema: a detailed review and analysis correcting for publication bias

Studies were included based on the following predefined criteria:

Studies were excluded if they focused primarily on outdoor fuel combustion, environmental tobacco smoke (ETS) as the main exposure, or were non-original research (e.g., reviews, conference abstracts, letters to the editor). Animal studies and studies with insufficient data for extraction were also excluded.

A comprehensive literature search was conducted in PubMed, Embase, Web of Science, and the Cochrane Library from their inception to May 28, 2025. The search strategy combined Medical Subject Headings (MeSH) and free-text terms related to both the exposure (e.g., “stove,” “cookstove,” “biomass,” “fuels,” “cooking,” “heating,” “coal,” “wood”) and the outcome (e.g., “Atopic Dermatitis,” “Atopic Eczema”). The detailed search strategy for each database is provided in. The reference lists of included articles were also manually screened to identify additional relevant studies. Supplementary Table S1

Two reviewers independently screened titles and abstracts to remove irrelevant articles, followed by a full-text review of potentially eligible studies to determine final inclusion. Any disagreements were resolved through discussion or consultation with a third reviewer.

Data were extracted independently by the same two reviewers using a standardized Excel spreadsheet. Extracted information included: first author, publication year, country, study design, sample size, participant characteristics (age), exposure type (fuel and usage), outcome definition and ascertainment method, and the most fully adjusted effect estimate with its 95% CI. When a single study reported multiple effect sizes for distinct, non-overlapping exposure subgroups (e.g., solid fuels vs. gas fuels, or cooking vs. heating) compared to a common reference group, each was treated as an independent effect size in our analysis. This approach allows for a more granular examination of different exposure scenarios. Consequently, 21 independent effect sizes were extracted from the 10 included studies.

The methodological quality of included studies was assessed independently by two reviewers. The Joanna Briggs Institute (JBI) Critical Appraisal Checklist for Analytical Cross-Sectional Studies was used for cross-sectional studies. This tool assesses eight domains, and studies with a score of ≥50% were considered for inclusion. For cohort studies, the Newcastle-Ottawa Scale (NOS) was used, which evaluates studies on three domains (selection, comparability, and outcome) with a maximum score of nine stars. Studies scoring seven or more stars were considered high quality. Detailed results of the quality assessment for each included study against each criterion are presented in. Supplementary Table S2

All statistical analyses were performed using the ‘meta’ package in R software (version 4.3.1). Odds ratios (ORs) and their 95% CIs were used as the primary effect measure as this was the metric reported by all included studies. It is important to note that while our analysis included one cohort study (9), it used multiple logistic regression analysis to control for confounders and thus also reported its findings as ORs. Therefore, no conversion or harmonization of different effect measures (e.g., from Relative Risk) was required. Statistical heterogeneity among studies was assessed using Cochran’s Q test (with a significance level of p < 0.10) and the I² statistic, where I2 > 50% was interpreted as substantial heterogeneity. A random-effects model (Der Simonian-Laird method) was used for the primary meta-analysis to account for anticipated between-study variance (10).

The initial database search yielded 939 articles. After removing 359 duplicates, 580 unique records remained. A review of titles and abstracts led to the exclusion of 510 articles that were clearly not relevant. The full texts of the remaining 70 articles were assessed for eligibility, of which 60 were excluded for reasons such as having an irrelevant exposure or outcome, being a review article, or lacking sufficient data. Ultimately, 10 studies met the inclusion criteria and were included in the meta-analysis. The PRISMA flow diagram detailing the study selection process is shown in Figure 1.

The 10 included observational studies were published between 1996 and 2024 and comprised a total sample size of 53,924 participants. The studies included nine cross-sectional designs and one prospective cohort study. They were geographically diverse, conducted in South Africa, China, Japan, Spain, Ukraine, Australia, Finland, and Germany. Most studies used the International Study of Asthma and Allergies in Childhood (ISAAC) questionnaire for outcome ascertainment, and all studies adjusted for basic confounders such as age and sex. The quality assessment indicated that the overall methodological quality of the included studies was acceptable; all cross-sectional studies had JBI scores above 50%, and the single cohort study was rated as high quality with a NOS score of 8 out of 9 stars. The detailed characteristics and quality assessments are presented in Tables 1, 2, respectively.

The random-effects meta-analysis, combining 21 effect sizes from the 10 studies, indicated a statistically significant positive association between indoor fuel use and the risk of AD. Several publications contributed more than one effect size to the analysis, as they assessed different types of indoor fuel exposures (e.g., solid fuels vs. gas) or different usage patterns (e.g., cooking vs. heating) separately. The forest plot in Figure 2 presents all 21 independent effect sizes extracted from the 10 included studies. The pooled OR was 1.158 (95% CI [1.051, 1.276], p = 0.0031). The analysis revealed moderate heterogeneity among the studies (I² = 43.0%, Cochran’s Q test p = 0.0195), supporting the decision to use a random-effects model. The forest plot summarizing the individual study results and the pooled estimate is shown in Figure 2.

Visual inspection of the funnel plot (Figure 3) revealed asymmetry, with a noticeable lack of studies in the bottom-left portion of the plot. This pattern suggests an absence of small studies reporting null or negative associations, a classic indicator of “small-study effects,” which may arise from publication bias.

This visual finding was corroborated by a comprehensive panel of statistical tests (Table 3). While Begg’s rank correlation test was borderline (p = 0.0533), Egger’s linear regression test showed highly significant asymmetry (p = 0.0006). This conclusion was strongly supported by Peters’ test (p < 0.0001), Deeks’ test (p < 0.0001), Macaskill’s test (p = 0.0006), and Tang’s test (p = 0.0006). The strong convergence of results from multiple tests, which rely on different statistical assumptions, provides robust evidence that the included literature likely represents a biased sample of all studies conducted on this topic.

To quantify the impact of the detected publication bias, two different correction methods were applied. The results, summarized in Table 4, demonstrate that the initial significant finding was not robust to bias adjustment.

First, the non-parametric trim-and-fill method estimated that nine studies were missing from the analysis due to bias. After imputing these hypothetical studies to create a more symmetric funnel plot (Figure 4), the adjusted pooled OR was reduced to 1.040 (95% CI [0.930, 1.164]), which is no longer statistically significant.

Second, parametric selection models, which explicitly model the probability of study publication, were employed. These models consistently attenuated the effect estimate to the null. The standard single-tailed selection model, for instance, yielded an adjusted OR of 1.000 (95% CI [0.992, 1.008]). The convergence of these methodologically distinct approaches provides strong evidence that the initial association was likely a statistical artifact driven by selective publication.

In contrast, the leave-one-out sensitivity analysis (Figure 5) showed that the initial uncorrected result was stable. Sequentially removing each study did not substantially alter the pooled OR (range: 1.13 to 1.18) or its statistical significance, indicating that the unadjusted finding was not unduly influenced by any single study.

Subgroup analyses were conducted to explore potential sources of heterogeneity (Table 5). The results are presented in a forest plot in Figure 6. When stratified by fuel type, a significant association was observed for solid fuels (OR = 1.13, 95% CI [1.01, 1.27]), but not for gas fuels (OR = 1.28, 95% CI [1.00, 1.65]). However, the test for a difference between these subgroups was not statistically significant (p = 0.670), indicating no clear evidence that the risk differs by fuel type.

When stratified by fuel usage, a significant association was found for cooking only (OR = 1.20, 95% CI [1.04, 1.38]), while no significant association was found for heating only or for combined heating and cooking. Similarly, the test for subgroup differences was not statistically significant (p = 0.900).

This systematic review and meta-analysis provides the first quantitative synthesis of the association between indoor fuel use and the risk of atopic dermatitis. The initial analysis of published literature suggested a modest, statistically significant 16% increase in the odds of AD associated with exposure to indoor fuel combustion (OR = 1.158). However, this finding proved to be fragile. A comprehensive and multi-pronged assessment revealed compelling evidence of publication bias, suggesting that studies with smaller or null effects were less likely to have been published. When this bias was statistically corrected using both non-parametric (trim-and-fill) and parametric (selection models) methods, the pooled effect estimate was attenuated to the null, and its statistical significance was completely eliminated. Therefore, the central conclusion of this study is that the weak positive association observed in the published literature is likely a statistical artifact, heavily influenced by the selective reporting of positive findings. Based on a rigorous, bias-adjusted analysis of the currently available evidence, it is not possible to conclude that indoor fuel use is an independent risk factor for AD.

The subgroup analyses yielded a counterintuitive result: no statistically significant difference in risk was detected between solid fuels and gas fuels, despite the well-established fact that solid fuel combustion releases far higher levels of harmful pollutants, particularly PM2.5. This finding does not necessarily imply that the risks are equivalent (15). Instead, it may reflect a profound methodological limitation common to the included primary studies: potential exposure misclassification (16).

While this bias is a concern in any observational design, it is especially pertinent here because the vast majority of studies that form our evidence base relied on questionnaires and long-term recall for exposure assessment. This retrospective assessment is highly susceptible to recall error, leading to non-differential exposure misclassification (15). A well-documented statistical consequence of such misclassification is a bias of the effect estimate (e.g., the OR) toward the null value of 1.0, potentially masking a true effect.

This reliance on recall is not a hypothetical concern but a consistent feature across the literature. For instance, studies by Zhang et al. (17), Vicedo-Cabrera et al. (18), Fedortsiv et al. (19), and the recent work by Bhuda et al. (4) all collected key exposure data on cooking fuels and home environments via parental questionnaires. Even earlier studies, such as Schäfer et al. (20), did the same. The inherent imprecision of asking a parent to recall details from months or years ago is a key mechanism that could lead to an underestimation of the true effect. Therefore, this measurement error, rooted in the design of the primary studies, serves as a core alternative explanation for our null finding after correcting for publication bias.

This type of imprecise measurement can lead to non-differential exposure misclassification, meaning that exposed and unexposed individuals are misclassified at similar rates. The well-documented statistical consequence of such misclassification in epidemiological studies is a bias of the effect estimate (e.g., the OR) toward the null value of 1.0. While this concept is most clearly demonstrated in cohort studies, it can also be a significant issue in cross-sectional designs that rely on long-term recall of exposure, potentially masking a true effect (21).

Beyond publication bias and exposure misclassification, several other factors could explain the null finding after bias correction. First, residual confounding is a major concern. Socioeconomic status, a strong determinant of both household fuel choice and health outcomes, may not have been adequately controlled for in all primary studies. Factors such as parental education, household income, and housing quality could confound the observed association, and the adjustments made in original studies might be insufficient. Second, co-exposure to other environmental pollutants was often not accounted for. Exposure to ambient air pollution or environmental tobacco smoke, which are often correlated with indoor fuel use, could independently affect AD risk and thus confound the results. Finally, potential misclassification of the outcome warrants consideration. The reliance on self-reported questionnaires like the ISAAC protocol, rather than a physician’s clinical diagnosis, in the majority of included studies could lead to non-differential misclassification of AD status, which would also tend to weaken any true association and bias the result toward the null.

The primary strength of this study lies in its comprehensive and rigorous approach to assessing and correcting for publication bias. By employing a wide array of statistical tests and multiple, methodologically distinct adjustment techniques, this analysis provides a robust evaluation of the stability of the published evidence. Further strengths include a systematic and reproducible search strategy, adherence to PRISMA reporting guidelines, and prospective protocol registration.

However, this review is subject to several limitations, most of which are inherited from the primary studies. First, the predominance of cross-sectional designs prevents any inference of causality, as it is impossible to establish whether the exposure preceded the onset of AD. Second, although moderate, the presence of statistical heterogeneity suggests that there are underlying differences between the studies that were not fully explained by the subgroup analyses. Third, the review was restricted to English-language publications, which may have introduced language bias. Finally, and most importantly, our conclusions are drawn from a body of literature constrained by the potential methodological weaknesses discussed above, including exposure misclassification, residual confounding, and outcome measurement error.

The findings of this meta-analysis highlight critical gaps and provide clear direction for future research in this area. To move the field forward and obtain a definitive answer, studies must overcome the methodological shortcomings of the past. The following recommendations are proposed:

By implementing these more rigorous designs, future research will be better positioned to accurately characterize the dose–response relationship and determine whether a true causal link exists between indoor fuel combustion and the risk of atopic dermatitis.

The original contributions presented in the study are included in the article/, further inquiries can be directed to the corresponding author. Supplementary material