Objectives

In this systematic review, our objectives were to 1) identify factors associated with developing CD from the published literature and 2) to combine data across studies to determine the pooled effect-size of the risk for most frequently reported factors.

Data sources & searches

Literature was searched from MEDLINE(OvidSP), EMBASE(OvidSP), Cochrane Database of Systematic Reviews and Cochrane Central Register of Controlled Trials (Cochrane Library, Wiley) from inception (1966 for MEDLINE, 1947 for EMBASE) to April 9, 2024. The search strategy included the broad key terms for the conditions of interest, for children and young people and for risk factors (Table A1 in S1 File). The initial search in May 2022 included keywords for 3 chronic conditions, the update in April 2023 and April 2024 focused on Coeliac Disease. No date or language limits were applied. A forward and backward citation search was conducted on included studies and relevant review articles using citationchaser [26]. All references were exported to Covidence for deduplication [27].

Study selection

Studies were selected for inclusion if they:

• Were observational studies or randomized trials.
• Reported an effect measure/an outcome from a statistical test for a risk/predictive factor for CD (see definition below).
• Had a comparison group without the factor of interest.
• Included data on children/young people aged 0–25 years.
• Were published in any country, language or years of publication.
• Reported on a predictive factor relevant to/possible to assess in a primary care context.

Studies were excluded if they:

• Reported risk/predictive factors not relevant to primary care (such as genetic factors).
• Focused on signs/symptoms only.
• Were case reports, case series and protocols.
• Were review articles; however, if they were topically relevant, they were tagged, and their references screened for inclusion.

Titles/abstracts and full texts were independently assessed by two reviewers for eligibility (CF and either AP or DS). Disagreements were resolved by discussion. At the stage of full text review, the specific reason for exclusion was documented.

Data collection & synthesis

Data were extracted from each report by CF and checked by a second reviewer (AP or DS). For studies with missing data, authors were contacted if the article was published within the past ten years.

For each report, data extracted included:

• The country/countries in which the study was conducted.
• The population size (total number of participants and number with CD).
• The eligible risk factor(s) investigated in the study.
• How each risk factor was defined/measured/categorized.
• The effect measure for each risk factor: odds ratio (OR), hazard ratio (HR), relative risk (RR), p-value if comparing proportions etc.
• Number of participants with and without risk factor in CD and non-CD group if needing to calculate an OR.
• Means and standard error of the mean (SEM) in CD and non-CD group (only applicable for height and weight).

Adjusted effect measures were selected over unadjusted if both were available. If a ratio estimate was not provided, 2x2 tables were constructed when possible and unadjusted ORs and 95% confidence intervals calculated.

Height and weight are continuous variables which cannot be appropriately compared through a ratio. For these variables, means and standard errors of the means for the CD and non-CD groups were extracted for use in calculating the standardized mean difference (SMD).

For study groups with multiple reports/published studies using data on the same study population, data were extracted for each report and then analysed together as a study group. If multiple reports included the same risk factor, one was selected to represent the study group. This choice was made hierarchically based on:

1. The availability of an adjusted effect measure.
2. The largest sample size of participants included in the analysis.
3. The consistency of the definition of the factor with other studies reporting on the same factor.
4. The recency of the publication.

CD was defined differently across studies depending on the year of publication and the availability of different types of testing. If only coeliac autoimmunity was assessed, this was used as an acceptable endpoint given the recent shift to diagnosis based on serology alone and not always requiring a biopsy in some jurisdictions. If data were presented on both coeliac autoimmunity and confirmed cases of CD for the same factor within a study, we extracted data based on the more specific definition of confirmed CD.

Table 1 provides the definitions used when extracting data for the factors included in meta-analysis. Extracted effect measures were rescaled to match these definitions to allow data to be combined across studies when necessary.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Definitions used when extracting data on each predictive factor for meta-analysis.

https://doi.org/10.1371/journal.pone.0306844.t001

Risk of bias assessment

The methodological quality and risk of bias of individual studies were assessed using the Newcastle-Ottawa Scale for cohort studies and for case-control studies [28]. A modified version of the Newcastle-Ottawa Scale was used for cross-sectional studies, and is included in the supplemental materials (Document A1 in S1 File) [29]. The scores were converted to Agency for Healthcare Research and Quality standards of good, fair or poor quality. Randomized trials were assessed using the Risk of Bias 2 (RoB 2) tool [30].

Eligibility for meta-analysis

Meta-analysis was conducted for factors reported in ≥ 5 studies. Age was excluded from meta-analysis as the way age was grouped and reported was inconsistent across studies so there were not at least three studies using similar definitions to be combined. Dietary Composition was similarly excluded as each study investigated different aspects of diet and dietary composition and could not be combined. Birth year was similarly excluded because studies investigated different years and could not be combined. T1DM was selected for meta-analysis post-hoc despite being reported in only four studies for face validity given its clinical relevance, inclusion in guidelines and known association with coeliac disease [8–11].

Meta-analyses

We calculated pooled OR estimates using log OR and their 95% confidence intervals (CIs) or standardized mean difference (SMD) for each factor with a random-effects meta-analysis using the Hartung-Knapp-Sidik-Jonkman (HKSJ) method [31]. Standardized mean difference was used for height and weight because the range of how these variables were reported across included studies (z-scores, percentiles and actual units of height and weight). Given the range of ratio effect measures reported in the included studies, ORs, HRs and RRs were all combined in the initial meta-analyses [32, 33].

Sensitivity analyses were conducted restricting to 1) only high or fair quality studies based on risk of bias assessment, 2) studies with >100 participants, 3) studies reporting odds ratios (as opposed to hazard ratios or relative risks) and 4) studies reporting adjusted effect measures. Sensitivity analysis was only performed if at least 3 studies could be combined for that factor. A sensitivity analysis was also conducted comparing the HKSJ random effects model to the DerSimonian and Laird (DL) random effects model. Statistical significance for all meta-analyses was assessed at a threshold of p-value under 0.05.

Reporting bias was assessed through funnel plot asymmetry. For height and weight (assessed through SMD and not pooled OR), Egger’s test was used to test for funnel plot asymmetry.

Given the nature of this review being predominantly observational trials, we elected not to use the Grading of Recommendations Assessment, Development, and Evaluation (GRADE) approach to assess certainty of evidence. Observational studies can provide the optimum evidence for the predictive value of an indicator, yet, in GRADE, observational studies start with a low level of quality [34, 35]. The comparison of the HKSJ and DL random effects models and sensitivity analyses restricting to adjusted measures only and excluding high risk of bias also contribute to an assessment of the certainty of evidence.

All analyses were conducted using STATA version 18 (Standard Edition, StataCorp, College Station, TX).

Patient & public involvement

We sought input from individuals with a range of experiences related to CD as parents and patients. Consultation took place at the stage of planning and drafting the protocol and throughout the analysis. The input from our patient and public partners was considered in study design and in plans for disseminating the study results. Speaking to parents and patients in the design stage informed the research question about the types of predictive factors focused on in this review and the decision to exclude symptoms which are already well understood and considered in the diagnostic process. Input from patients and parents at the analysis and write-up stage informed the clinical implications section of the discussion.

Study selection & characteristics

Our database searches and the forward/backward citation search of all included records and relevant reviews identified 11575 records after de-duplication. We selected 410 records for full text review, of which 183 fulfilled inclusion criteria (see Fig 1). Included articles were published from 1983–2024.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Flowchart for eligible studies identification according to PRISMA 2020 guidelines.

PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses).

https://doi.org/10.1371/journal.pone.0306844.g001

The summary characteristics of all included studies (by study group), including the predictive factors reported in each, are presented in full in Table 2, and the references are included in Table A2 in S1 File. Of the 183 reports, 8 were randomized trials, 48 were case-control studies, 119 were cohort studies and 8 were cross-sectional. 102 reports were unique studies and the remainder presented data from 13 study groups which were analysed together, giving a total of 115 studies. Reports presented data from a range of countries (number of reports given in parentheses): Sweden (43), International cohorts (23), Italy (21), USA (15), Norway (13), Israel (12), Denmark (7), UK (7), Finland (6), Netherlands (4), Spain (5), Turkey (5), Germany (4), Iran (2), Jordan (2), Switzerland (2), Saudi Arabia (2), Cyprus (1), France (1), Greece (1), UAE (1), India (1), Poland (1), Oman (1), Brazil (1), Ethiopia (1), and Pakistan (1).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Characteristics of eligible studies.

https://doi.org/10.1371/journal.pone.0306844.t002

We identified 145 predictive factors, of which 30 were reported in ≥5 studies. Excluding age, dietary composition and birth year and adding T1DM (as discussed in the methods), 28 factors were included in the meta-analysis. The frequency with which each predictive factor was reported by a unique study/study group is presented in Table 3 for those reported in ≥5 studies and in Table A3a/A3b in the S1 File for those reported in ≤4 studies. Table A4 in the S1 File summarizes which studies reported a predictive factor as significant and which as non-significant for all risk factors reported in ≥5 studies. Age was reported in seven studies and the extracted data is presented in Table A5 in S1 File given its potential relevance, but given the heterogeneity of definitions in the individual studies, it was not possible to include age in the meta-analysis.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Predictive factors for a diagnosis of coeliac disease, and the frequency in which they were reported in the included studies.

This table is restricted to factors reported in ≥5 studies and therefore eligible for inclusion in the meta-analysis. Some studies did not report sufficient data to allow for inclusion in meta-analysis, so the total studies figures included in this table maybe higher than the number of included studies for the corresponding meta-analysis.

https://doi.org/10.1371/journal.pone.0306844.t003

We excluded studies which did not include a comparator group, which meant we were not able to include studies which suggested potential associations. For example, this led to the exclusion of studies reporting high rates of CD amongst children with Trisomy 21 or T1DM due to a lack of a comparison group without the condition [36, 37].

Risk of bias

The study quality/risk of bias results are summarized in Table 4 and the score breakdown and justification is available in the S2 File. Most studies were rated as good quality, with only eight of fair quality or having some risk of bias, and one study considered poor quality.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Study quality and risk of bias assessment for all included studies, assessed using the Newcastle Ottawa Scale (NOS) for cohort studies or Case-Control (C-C) studies, the modified NOS for Cross-Sectional (C-S) studies, or the Risk of Bias 2 (RoB 2) Tool for Randomized Trials.

https://doi.org/10.1371/journal.pone.0306844.t004

Results of meta-analyses

Meta-analyses of the individual risk factors are presented in Table 5 and summary forest plot Fig 2, individual forest plots are in Figures A1-A5 in S1 File. The following factors were statistically significant: having type 1 diabetes (OR 8.70, 95% CI 7.70–9.83, I² = 0.0%), having a first degree relative with coeliac disease (OR 5.19, 95%CI 2.48–10.86, I² = 96.3%), being of white ethnicity (OR 2.56, 95%CI 1.40–4.70, I² = 55.5%), having thyroid disease (OR 2.16, 95%CI 1.61–2.90, I² = 0.0%), being female (OR 1.53, 95%CI 1.39–1.69, I² = 74.6%), antibiotic use in early childhood (OR 1.21, 95%CI 1.04–1.42, I² = 64.9%), having frequent respiratory infections in early childhood (OR 1.47, 95%CI 1.03–2.09, I² = 78.7%), and being born in the summer (OR 1.09, 95%CI 1.00–1.19, I² = 67.5%).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Summary forest plot for predictive factors included in meta-analysis.

Upper region of plot displays pooled odds ratio estimates and lower region of plot displays standardized mean difference.

https://doi.org/10.1371/journal.pone.0306844.g002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 5. Sensitivity analysis of meta-analysis for each predictive factor comparing Hartung-Knapp-Sidik-Jonkman (HKSJ) and DerSimonian and Laird (DL) odds ratio (OR) estimates or standardized mean difference (SMD) and if they were significant for each predictive factor.

Significance assessed as p≤0.05.

https://doi.org/10.1371/journal.pone.0306844.t005

Sensitivity analysis comparing the HKSJ and DL random effects models (Table 4) found six additional predictive factors to be significant when using DL but not with HKSJ. These were: older age at T1DM diagnosis (protective, OR 0.64, 95%CI 0.47–0.88, p = 0.005, I² = 84.6%), breast feeding (protective, OR 0.79, 95%CI 0.72–0.88, p<0.001, I² = 94.7%), more frequent gastroenteritis (OR 1.48, 95%CI 1.13–1.92, p = 0.004, I² = 78.8%), greater gluten ingestion (OR 1.25, 95%CI 1.06–1.48, p = 0.008, I² = 91.2%), and more frequent infections (OR 1.22, 95%CI 1.06–1.39, p = 0.005, I² = 77.0%). A heavier weight was associated with a lower risk of CD (SMD -0.21, 95%CI -0.41 to -0.02, p = 0.029, I² = 60.2%) representing a small effect size [38].

The remaining factors were all found to be not significant in their respective meta-analyses including in sensitivity analysis comparing HKSJ and DL. Breast feeding at gluten introduction, vitamin D supplementation, higher birth weight, and higher maternal education all had point OR estimates suggesting a possible trend but wide confidence intervals encompassing the line of no effect (OR of 1). Higher income, higher BMI, maternal smoking, premature gestational age, being born via c-section, higher maternal age, multi-parity, age at gluten introduction and having ≥1 older sibling were all found to be non-significant and had pooled OR estimates 0.9–1.1. Taller height had a point estimate of SMD similar to weight which would show a lower risk of CD, but the confidence interval included zero.

Sensitivity analysis restricting to studies with >100 participants did not change the results (results not shown). Excluding hazard ratios and relative risks made summer season of birth not significant with a confidence interval encompassing 1, and did not meaningfully change the results for any other predictive factor (including those with pooled OR estimates >2). Restricting to only adjusted measures likewise did not change the results for those factors with ≥3 studies remaining to proceed with meta-analysis. As most studies were of good or fair quality, restricting analysis to only studies of good quality did not change the results for those factors with ≥3 studies remaining to proceed with meta-analysis. No studies considered to be of poor quality/high risk of bias were included in the main meta-analyses.

Reporting biases

Funnel plots are shown in Figures A6-A10 in S1 File. For height and weight, Egger’s test was not significant. Most predictive factors had reasonably symmetric funnel plots except for first degree relative with CD, infections, gluten ingestion, age at T1DM diagnosis, respiratory infections, and antibiotics suggestive of potential publication bias or selective outcome reporting for these factors.

Principal findings

To our knowledge, this is the only systematic review of studies examining the association between predictive factors relevant to clinicians seeing initial presentation of children and the risk of CD specifically in children and young people. This systematic review compiled evidence from the peer-reviewed literature to identify 145 factors assessed for an association with a diagnosis of CD. Meta-analyses of the 28 most reported predictive factors revealed 14 factors to be significant on one or both of the random effects models and quantified the strength of these associations. The remaining 14 factors were not significantly associated with CD using either model. Of the factors found to be significant, five had OR estimates >1.5, and the remaining all had smaller effect sizes.

Comparison with previous literature

Individual predictive factors have been summarized in other reviews, such as the significance of breast feeding [39–46], infant feeding [47], perinatal factors [48] or gluten introduction [49]. Some reviews have approached the question of predictive factors for CD more broadly [15, 50, 51]. A comprehensive systematic review and meta-analysis was recently published by Elwenspoek et al. on the “diagnostic indicators for coeliac disease” [15]. This review included both adults and children and included symptoms [15]. Our review differs in that it excluded symptoms and adds to the work of Elwenspoek et al. by including a broader array of potential predictive factors, such as perinatal risk factors and age at gluten introduction). Both their review and this one restricted meta-analysis to factors with ≥5 studies. The results are consistent between our meta-analysis previous reviews in the significant association of T1DM, thyroid disease and family history of CD [15, 52–54].

Family history of CD and autoimmune conditions are well-documented risk factor given the genetic/heritable factors associated with CD [55], consistent with our findings for family history, T1DM and thyroid disease as predictive factors for CD [55]. These are all early indicators which could be used as criteria for screening. Multiple autoimmune conditions are more common in female patients, which is consistent with our finding of being female as a predictive factor [56]. The effect of ethnicity may be explained through the prevalence of genetic factors in certain populations, and by other non-genetic factors such as by socioeconomic status or day care attendance as shown in other studies, so it needs to be considered in context [57].

This analysis has found that early life antibiotic exposure and respiratory infections are both associated with CD, and having more infections was found to be significantly associated using the DL model. Some reports have suggested mechanisms for how antibiotics and infections could increase the risk of CD [58, 59]. The reports varied in how antibiotic use and respiratory infections were defined but they all focused on early childhood. The proposed mechanisms relate to the impact of infections and antibiotics on the composition of the gut microbiome, the regulation of the immune system by certain infections and how some bacterial peptides may mimic immunogenic gluten peptides, priming the immune system to respond to gluten [58, 59]. This impact of infections may also explain why season/month of birth is associated with CD in this analysis, as children born in the summer experience their first viral season/infections at a different age than children born in the winter [60].

Six additional factors (older age at T1DM diagnosis, being breast fed, more frequent infections, increased amount of gluten ingestion, more frequent gastroenteritis, and heavier weight) had significant associations in sensitivity analyses using the DL random effects models. Some of the primary studies included in the meta-analyses reported significant associations [20, 21, 61–82]. Given their lack of significance in the primary analysis using the HKSJ random effects model, there remains uncertainty about the strength of these associations, and these factors may be candidates for further research to address this uncertainty. Lower height and/or weight may be manifestations of CD rather than early indicators and may suggest late diagnosis or failure to thrive. However some evidence suggests growth rate may be affected before any other symptoms or disease manifestations are present so in some children may be a useful indicator [83].

There is uncertainty in the evidence as to whether breastfeeding is protective against CD or if it delays onset or changes the symptom profile at presentation. Breastfeeding has been the subject of previous reviews [43, 84]. One paper included in the meta-analysis for breastfeeding found a higher rate of CD amongst children breastfed for ≥4 months [20]. This study only included children with T1DM, and we had to calculate an unadjusted odds ratio from the raw prevalence data included in the paper, so it is likely that this association may be explained by other factors as all other studies found breastfeeding to be protective or to have no significant effect.

Strengths & limitations

Our review extends what is known by compiling evidence in a systematic review and meta-analysiss, identifying predictive factors and strengths of associations. Our work aligns with the James Lind Alliance and Coeliac UK research priorities by summarising the current knowledge on the risk factors for the development of CD, and by helping to better understand the association between CD and other conditions [23].

We utilised a comprehensive search strategy including forward and backward citation searching of included studies and relevant reviews to capture all published articles relating to the study question. Screening and data extraction was performed in duplicate and strict inclusion criteria were applied to minimize bias.

The clinical and methodological diversity between studies can be observed in the reported I² statistics which are often indicative of moderate to high heterogeneity [85]. The direction of associations was typically aligned across studies for the same predictive factor despite the heterogeneity. We anticipated there may be heterogeneity and, therefore, chose to use a random effects model [85]. We chose the HKSJ random effects model as our primary form of meta-analysis which performs better in simulation models than DL and is less likely to overestimate possible effects given the wider confidence intervals [31].

As demonstrated by the instances of funnel plot asymmetry, there is a possibility of some publication and/or selective outcome reporting bias. We were limited in our ability to perform subgroup analyses (such as by sex) because insufficient studies reported stratified data. Sensitivity analyses restricting to larger study size (>100 participants) and fair to high quality studies were performed when possible and they did not change the results from the primary analyses which offers confidence in the results of these meta-analyses.

Differences in definitions and cut-offs of some predictive factors used in the primary studies could have contributed to the heterogeneity in study results in the meta-analysis and may have limited the ability to fully evaluate the impact of some factors. For example, breast feeding was assessed across studies from between 30 days of breast feeding to >12 months. To combine these studies together, we simplified the definition to a dichotomous variable of ‘any breast feeding’ compared to ‘no breast feeding’. Some studies included routinely collected data which presents the limitation of the likely under-ascertainment of diagnosis. The cohorts and cases/controls used in many studies recruited people with known diagnoses or specific risk factors or from specific clinical settings which can introduce a bias about the applicability of those results to the larger population.

We combined OR, HR, and RR together in the meta-analyses given the range of effect measures reported in the primary studies. The difference between these effect measures is minimal when the outcome is rare, as was the case with the factors included in the meta-analyses, but is more pronounced when the association is greater/when the event is common [32, 33]. Including hazard ratios and relative risks in calculating a combined odds ratio potentially underestimates the overall effect as odds ratios will exceed both relative risks and hazard ratios when greater than one [33]. We also combined adjusted and unadjusted effect measures (preferring adjusted measures when they were available) [32]. Sensitivity analysis restricting to only OR and to only adjusted measures were performed when possible and did not change the overall conclusions. Although there is a risk of bias, the results of our sensitivity analysis provide confidence in the reported findings.

Clinical implications

Predictive factors can be included in clinical guidelines to guide testing and clinical decision making. Clinical guidelines on CD typically focus on signs and symptoms, sometimes alongside predictive factors/associated conditions, such as T1DM, thyroid disease and having a first degree relative with CD [8–11]. These factors were all found to have significant and strong associations in this review, adding to the evidence supporting these guidelines.

Our meta-analyses have not identified other factors which could be easily used to risk stratify in isolation but have identified factors which may be suitable candidates for future research. The other factors identified in this review as significant (being female, being of white ethnicity, having been breast fed, history of early childhood antibiotic use and frequent/severe infections etc.), are easy to assess in an initial clinical assessment and may have greater value as predictive factors in the context of specific symptoms or when combined. Future studies could investigate the feasibility and predictive value of combining candidate factors which can be assessed in primary care consultation with signs and symptoms to aid in risk stratification. These factors and symptoms together could be used to create clinical risk prediction tools or automated clinical prompts for use in primary care. Children with CD can experience a range of non-classical CD symptoms [86]; these children are more likely to experience delayed diagnosis, so factors which aid in identifying at-risk children prior to diagnosis would be of particular benefit to this population [87–89].

Diagnostic delay for CD in children may range from months to years from symptom onset with an average delay of 5 months; there is evidence that when diagnosis is delayed longer than three years, children have lower body weight and shorter stature at the time of diagnosis [89, 90]. Previous misdiagnosis and certain symptom patterns associated with longer delays and female children and children at older ages were also more likely to have a delayed diagnosis [90]. These findings highlight an opportunity for further research to identify children with patterns of symptoms and predictive factors at risk for longer delays in diagnosis for targeted testing.

Ultimately, the value of understanding and utilising predictive factors for CD lies in the utility of knowing which factors are and which are not associated with CD. This can guide clinicians as to what specific information is useful to assess in a time-limited setting and to help them to identify children at increased risk of CD and the need for testing.

Some of the significant predictive factors included in our meta-analyses are modifiable risk factors (breast feeding, antibiotic use, and gluten ingestion). Although these cannot be modified for an individual patient at the time of screening, these can be considered along with other evidence to support recommendations and guidance regarding advice to parents on breast feeding and diet, and antibiotic prescribing in early childhood.