Demographic differences

Important demographic considerations in GMB studies are sex and age. Substantial differences in the gut microbiota are attributable to sex differences in mammals (Org et al., Reference Org, Mehrabian, Parks, Shipkova, Liu, Drake and Lusis2016; Kim et al., Reference Kim, Unno, Kim and Park2020). Because of this, any study which includes one sex in one arm and a different sex in another should be classified as having a high ROB. In addition to the ROB from sex, other demographic factors may also introduce confounding bias into the studies being examined. The GMB changes with age across numerous conditions, disease models, and species impacting microbial diversity and biome composition (Ticinesi et al., Reference Ticinesi, Nouvenne, Cerundolo, Catania, Prati, Tana and Meschi2019; Liu et al., Reference Liu, Lv, Wang, Yang, Li and Qian2020). Therefore, age differences between cohorts and study arms should be assessed. If the study being examined uses organisms of one age in one arm and a different age in a second arm, it should be classified as having a high ROB. The age gap which introduces significant confounding bias, varies by organism. An example of an age gap that would introduce a high ROB is 8-week-old mice versus 1-year-old mice (Yoon et al., Reference Yoon, Ryu, Moon and Kim2021).

Habitat stability

The habitat in which organisms are kept substantially impacts their GMB (Singh et al., Reference Singh, Brass, Cruickshank and Knight2021). Mice, common subjects of microbiome research, are known to have highly variable microbiomes on arrival at a facility, likely because of transportation stress on the microbiome itself and the immune system and hormonal functions of the host organism (Capdevila et al., Reference Capdevila, Giral, Ruiz de la Torre, Russell and Kramer2007; Montonye et al., Reference Montonye, Ericsson, Busi, Lutz, Wardwell and Franklin2018; Lipinski et al., Reference Lipinski, Zhou, Gurczynski, Erb-Downward, Dickson, Huffnagle, Moore and O’Dwyer2021). Studies that do not allow for microbiome stabilisation before research begins risk confounding bias due to a lack of habitat stability. Organisms should be acclimated to the study condition before baseline measurements or interventions are performed. However, an extensive acclimation period risks microbiome drift occurring due to the increasing age of the organism or other unknown factors, so habitat stabilisation must be time-limited (Hoy et al., Reference Hoy, Bik, Lawley, Holmes, Monack, Theriot and Relman2015). Additional bias would also be introduced if the acclimation period is included in the interventional period of the research.

Genotype, familial, and source differences

Subject genotype, degree of familial relation, and in the case of animal models, the source can significantly impact GMB composition. Differences in the genotype of animal models have been found to impact the diversity and abundance of organisms (Campbell et al., Reference Campbell, Foster, Vishnivetskaya, Campbell, Yang, Wymore, Palumbo, Chesler and Podar2012; McKnite et al., Reference McKnite, Perez-Munoz, Lu, Williams, Brewer, Andreux, Bastiaansen, Wang, Kachman, Auwerx, Williams, Benson, Peterson and Ciobanu2012; Leamy et al., Reference Leamy, Kelly, Nietfeldt, Legge, Ma, Hua, Sinha, Peterson, Walter, Benson and Pomp2014). For this reason, if the study being evaluated uses organisms of significantly different genotypes, such as the use of different strains of mice from the Collaborative Cross, where the effect of genotype difference is not the target of the study, it should be classified as having a high ROB. Suppose the study uses a similar genotype between treatment groups, such as the same strain of inbred animal model or monozygotic twin subjects. In that case, it should be considered a low ROB for confounding due to the genotype effect.

Regarding familial relation, genetically related subjects have been demonstrated to share a core of similar GMB for up to three generations in the female line (Turnbaugh et al., Reference Turnbaugh, Hamady, Yatsunenko, Cantarel, Duncan, Ley, Sogin, Jones, Roe, Affourtit, Egholm, Henrissat, Heath, Knight and Gordon2008; Valles-Colomer et al., Reference Valles-Colomer, Bacigalupe, Vieira-Silva, Suzuki, Darzi, Tito, Yamada, Segata, Raes and Falony2021). With animal models, breeding within familial relations is often used to maintain genotypically and GMB homogeneity (Hufeldt et al., Reference Hufeldt, Nielsen, Vogensen, Midtvedt and Hansen2010). A caution regarding inbreeding is that while selective breeding between siblings can create a more stable and uniform GMB composition, the effects of genetic drift can also introduce confounders across multiple generations that may affect experimental reproducibility with subsequent generations (Laukens et al., Reference Laukens, Brinkman, Raes, De Vos and Vandenabeele2016).

Additionally, with animal models, an organism’s litter of origin impacts the gut microbiota (Fujiwara et al., Reference Fujiwara, Watanabe and Sonoyama2008; Vilson et al., Reference Vilson, Ramadan, Li, Hedhammar, Reynolds, Spears, Labuda, Pelker, Björkstén, Dicksved and Hansson-Hamlin2018). This may relate not only to parent genetics but also to the host of maternal factors that can affect the development of progeny GMB, including mode of delivery, maternal diet, maternal stress, and maternal antibiotic use (Bailey et al., Reference Bailey, Lubach and Coe2004; Friswell et al., Reference Friswell, Gika, Stratford, Theodoridis, Telfer, Wilson and McBain2010; Stokholm et al., Reference Stokholm, Schjørring, Eskildsen, Pedersen, Bischoff, Følsgaard, Carson, Chawes, Bønnelykke, Mølgaard, Jacobsson, Krogfelt and Bisgaard2014; Golubeva et al., Reference Golubeva, Crampton, Desbonnet, Edge, ’Sullivan O, Lomasney, Zhdanov, Crispie, Moloney, Borre, Cotter, Hyland, O’Halloran, Dinan, O’Keeffe and Cryan2015; Walker et al., Reference Walker, Clemente, Peter and Loos2017; Zhang et al., Reference Zhang, Li, Chen, Lee, Gao, Yuan, Nie and Sun2021). For these reasons, if the study being examined utilises organisms from differing litters (from separate mothers or separate deliveries from the same mother) that have not yet reached their mature adult development and are not randomly assorted between research arms, it should be classified as having a high ROB. Suppose a study uses organisms from the same mother and litter or randomly assorts progeny from different mothers and litters. In that case, it should be classified as having a low ROB.

Regarding sourcing of animal models, subjects sourced from different vendors have substantial differences in GMB at baseline (Rasmussen et al., Reference Rasmussen, de Vries, Kot, Hansen, Castro-Mejía, Vogensen, Hansen and Nielsen2019; Wolff et al., Reference Wolff, Jacobs, Haak, Roelofs, de Vos, Hugenholtz and Wiersinga2020; Long et al., Reference Long, Svenson, Mourino, Michaud, Fahey, Waterman, Vandegrift and Adams2021). The microbiological or physiological basis of these effects is unknown but may be due to differential exposures to environmental or infectious factors between vendors (Mandal et al., Reference Mandal, Denny, Waide, Li, Bhutiani, Anderson, Baby, Jala, Egilmez and Schmidt2020).

Extreme diet

Dietary differences have been shown to alter the abundance of most gut microbes (Daniel et al., Reference Daniel, Gholami, Berry, Desmarchelier, Hahne, Loh, Mondot, Lepage, Rothballer, Walker, Böhm, Wenning, Wagner, Blaut, Schmitt-Kopplin, Kuster, Haller and Clavel2014; Do et al., Reference Do, Lee, Oh, Kim and Park2018; Ang et al., Reference Ang, Alexander, Newman, Tian, Cai, Upadhyay, Turnbaugh, Verdin, Hall, Leibel, Ravussin, Rosenbaum, Patterson and Turnbaugh2020; Li et al., Reference Li, Yang, Zhang, Jiang, Zhang, Wang, Gong, Zhao and Zhang2021). Because of this, maintaining the diet of interest is essential to avoid introducing confounding bias to the study. However, it may not always be possible to strictly control diet. This is especially relevant to clinical studies involving humans. In this situation, an evaluation of bias must note how a study documented these diet variations.

GMB normalisation

It is important to assure organisms being studied in research have similar baseline GMB. This allows for more definitive inference as to the effect of the intervention. Several strategies have been used to make the GMB as similar as possible over time. Removal of the entire GMB through the use of germ-free mice can allow for the artificial seeding of a select group of organisms (Yi and Li, Reference Yi and Li2012; Kennedy et al., Reference Kennedy, King and Baldridge2018). However, the use of these mice necessarily limits the generalizability of a study. For this reason, research often uses organisms with populated GMBs and relies instead on antibiotics to homogenise the microbiome. The use of antibiotics introduces additional risks of bias which must be considered when evaluating a study (Theriot et al., Reference Theriot, Bowman and Young2016). The most significant ROB arises from beginning the intervention of interest before the gut microbiota has stabilised after normalisation with antibiotics. The GMB continues to fluctuate unpredictably for long periods following antibiotic administration (Merenstein et al., Reference Merenstein, Fraser, Roberts, Liu, Grant-Beurmann, Tan, Smith, Cronin, Martin, Sanders, Lucan and Kane2021). This variance has been found for at least a year after antibiotic usage in humans and for times ranging between 1 week and 16 weeks in mice depending on the length of the course of antibiotics used (Rashid et al., Reference Rashid, Zaura, Buijs, Keijser, Crielaard, Nord and Weintraub2015; Elvers et al., Reference Elvers, Wilson, Hammond, Duncan, Huntley, Hay and van der Werf2020; Zhu et al., Reference Zhu, Fu, Dong, Guo, Cao and Wu2021). However, short, or single doses of antibiotics such as those often used to normalise the microbiome allow for substantial stabilisation of the GMB within 7 days (Gu et al., Reference Gu, Gong, Zhang, Chen, Wu, Xu, Fang, Wang and Tang2020).

A third method used to standardise the GMB is to intermix the bedding of multiple cages and then redistribute it (Miyoshi et al., Reference Miyoshi, Leone, Nobutani, Musch, Martinez-Guryn, Wang, Miyoshi, Bobe, Eren and Chang2018). This method is less invasive than antibiotic usage and has a lower risk of long-term impact on the GMB than the use of antibiotics. The use of homogenisation of the bedding allows for similar microbiomes to develop in more mice than can be practically housed in a single cage, where the organisms also share all of their bedding (McCafferty et al., Reference McCafferty, Mühlbauer, Gharaibeh, Arthur, Perez-Chanona, Sha, Jobin and Fodor2013).

Because of the impact of different methods of GMB normalisation, it is critical to note the method that was used to normalise the GMB and how long before the intervention this normalisation was completed.

Extreme genotype

Host genotype shows a stable and heritable impact on GMB composition (Goodrich et al., Reference Goodrich, Davenport, Beaumont, Jackson, Knight, Ober, Spector, Bell, Clark and Ley2016). In the context of GMB research, extreme genotype selection refers to the selection of GMB subjects with genotypes that vary significantly between subjects within a study. Selection of subjects with identical or similar genetic makeup limits genotype confounding effects. A subject with an established history of use along with maximised genetic correlation can be considered a low risk of selection bias. For example, while inbred Balb/C mice do have an extreme genotype, they also have a long-established history of use in immune modulation studies with their known Th2 immune response wherein they exhibit low IFNy and high IL-4 production (Mills et al., Reference Mills, Kincaid, Alt, Heilman and Hill2000; Watanabe et al., Reference Watanabe, Numata, Ito, Takagi and Matsukawa2004; Khan et al., Reference Khan, Rehman, Farooqui, Siddiqui, Ayub, Ramzan, Wang and Xin2022). Furthermore, prior literature has established the correlation between subject genetics and variation in the GMB population and subsequent disease states (Xu et al., Reference Xu, Fu, Sun, Jiang, Miao, Shuai, Gou, Ling, Yang, Wang, Chen and Zheng2020).

Randomisation or demographic balancing sufficiently applied

Randomisation is essential in ensuring subject-level differences between participants in the intervention and control groups can be attributed to chance alone. It is a standard method that attempts to create the necessary pre-intervention equivalence between groups, allowing for conclusions based on the effect of the intervention. In trials where randomisation was not appropriately utilised, the outcome was overestimated by up to 40% compared to trials where randomisation was utilised (Suresh, Reference Suresh2011). If randomisation was not applied, implementing demographic balancing is an appropriate measure to ensure adequate control and intervention arms distribution. Any demographic balancing performed should be sufficiently described in the study. This method focuses on ensuring each group is demographically balanced at baseline to lessen the difference between groups and utilise randomisation if no subject background information is available (Saint, Reference Saint2015). Both randomisation and demographic balancing can be applied to human and animal model studies. For example, in studies utilising syngeneic mice, randomisation must be performed outside the scope of human intervention in that random number generators should assign mice numbers which can then correlate to intervention and control groups, hence this places randomisation outside the scope of human influence, limiting bias to a maximum degree. In syngeneic animals, demographic balancing would have a limited impact on the bias, however, wherein studies utilise genetically unrelated animals, the need for implementation of both randomisation and demographic balancing is necessary for limiting substantial bias (Hirst et al., Reference Hirst, Howick, Aronson, Roberts, Perera, Koshiaris and Heneghan2014). Similar principles apply in human studies. Given a majority of human studies utilise genetically unrelated subjects, randomisation is required to avoid the high ROB. In human studies, a step beyond randomisation should be taken, that is, implementing blinded randomisation with a description of the randomisation protocol to give the reader the ability to discern breaks in randomisation or similar bias control methods within the study (Chalmers et al., Reference Chalmers, Celano, Sacks and Smith1983).

Intervention bias

Bias in intervention can occur when interventions or outcomes are inappropriately selected for or measured. In non-differential misclassification, test subjects’ exposures are misidentified, and they are categorised into the wrong group (McCoy, Reference McCoy2017). This misclassification can dilute the effect of the intervention causing effect estimates to favour the null (LaMorfe, Reference LaMorfe2016). The probability of non-differential misclassification is equal across all groups. Bias may be reduced by ensuring a proper background check on test subjects and equalising any differences. On the other hand, differential misclassification occurs when the misclassification of exposure or outcome is not equal between subjects and is less easily predictable in whether it will bias results towards or away from the null. Therefore, the probability of assigning subjects to the wrong group differs based on the individual. This may also introduce recall bias towards recalling specific exposures because the subject has the disease state versus a subject that does not. In GMB studies, this may present in the form of researchers explaining results that show a significant effect as attributed to specific causes but leaving out explanations for non-significant results. Because this type of misclassification is more applicable in case studies, it is less relevant for animal studies but can be prominent in human studies (Spencer et al., Reference Spencer, Mahtani, Brassey and Heneghan2018).

Validation of method

The establishment of an effective intervention is imperative for a successful study. Before the experiment, researchers must verify that their chosen intervention method will produce the intended effect. In studies where this is not done, the produced results may or may not be relied on because the protocol was never validated. Verification can be internal (tested and proved by the researchers) or external (via other established studies). If the study calls for a particular disease state to be expressed, it must be validated that the test subjects have the disease state. In studies that call for a specific procedure, there can be potential bias in how the readers know the procedure was correctly obtained if it is not reported. For example, in microbiome hypertension studies, animal subjects were tested based on blood pressure measurements by a well-established method, tail-cuff plethysmography (Marques et al., Reference Marques, Jama, Tsyganov, Gill, Rhys-Jones, Muralitharan, Muir, Holmes and Mackay2019). If a lesser-known and validated method was used, it could introduce a high ROB if researchers did not verify that their method was accurate. When testing for the effect of a disease state as influenced by the microbiome, it is helpful to transplant the experimental group microbiome into a germ-free animal model to confirm the effect. This reduces an intermediate ROB by demonstrating that the effect of the intervention is associated with the levels of change in the microbiome (Gottfredson et al., Reference Gottfredson, Cook, Gardner, Gorman-Smith, Howe, Sandler and Zafft2015).

Cause/category of missing data

Missing data falls into multiple categories based on the mechanism of missingness: Missing Completely at Random (MCAR), Missing at Random (MAR), and Missing Not at Random (MNAR) (Groenwold and Dekkers, Reference Groenwold and Dekkers2020). These categories apply assumptions to missing data based on the cause. MCAR assumes that data is missing due to a factor entirely unrelated to the study. MAR assumes data is missing due to observed variables relevant to the study. MNAR assumes data is missing based on unknown or not quantifiable variables to the authors. MAR and MNAR are most relevant to clinical research, specifically in regard to patient drop-out, including clinical GMB trials (Pugh et al., Reference Pugh, Brown and Enserro2021). Sampling zeroes in microbiome data are a more generalised form of missing data but are primarily reminiscent of MAR (Kaul et al., Reference Kaul, Davidov and Peddada2017a, Reference Kaul, Mandal, Davidov and Peddada2017b). Each of these areas will be further discussed in the following sections. Under MAR, studies may utilise various statistical imputation techniques to replace missing data, though the most well-known and effective method is multiple imputations (Spineli et al., Reference Spineli, Fleming and Pandis2015). With MNAR, various statistical modelling techniques may address missing data. Such techniques are further discussed in relation to GMB studies in section “Sequencing depth and sampling zeroes.” The distinction between MAR and MNAR also indicates whether bias related to missing data is entirely removable in analysis – the former can, while the latter cannot (Mack et al., Reference Mack, Su and Westreich2018). This should not be confused with the notion that MNAR assumptions immediately denote a study as biased. If the missingness in MNAR or MAR is independent of the outcome, then the study may be unbiased in regard to missing data. Thus, a study with MNAR data is not necessarily high ROB.

Notably, a significant number of studies do not clearly state the mechanism of missingness or adjust for missing data (Carpenter and Smuk, Reference Carpenter and Smuk2021). It is important that studies distinguish the mechanism of missingness or explain relevant missing data. If a study does not acknowledge missingness in data or ensures the absence of missing data, the study may be considered high ROB. If a study acknowledges missing data but does not adequately address it through MAR/MNAR distinction and proper statistical techniques related to its missing data category, then the study may be considered intermediate ROB. If a study demonstrates all of this, it may be considered low ROB.

Subject drop-out

Missing data in the form of patient drop-out has a marked effect on statistical power, type 1 error, and various outcome measures (Thompson et al., Reference Thompson, Berk, Dean, Kohlmann, Jeavons, Bush and Copolov2011; Fiero et al., Reference Fiero, Huang, Oren and Bell2016; Cai et al., Reference Cai, Gewandter, He, Turk, Dworkin and McDermott2020). In traditional clinical research, missing data has a clear effect on useful measures, such as relative risk and risk ratio calculations. Further, although researchers attempt to minimise drop-out and its statistical effects, drop-out ratios were reported to be greater than 40% depending on the study and the degree of unpleasantness in medical interventions to the patient (Schnicker et al., Reference Schnicker, Hiller and Legenbauer2013; Li et al., Reference Li, Pei, Liu and Zhang2021). Consequently, it has been proposed that a 20% drop-out ratio is reasonable (Furlan et al., Reference Furlan, Pennick, Bombardier, van Tulder, Board and Group2009; Cramer et al., Reference Cramer, Haller, Dobos and Lauche2016). Interestingly, it has been shown that faecal sampling of patients in GMB studies has not been a significant reason for drop-out, suggesting typical sources of patient non-retention (Vandeputte et al., Reference Vandeputte, Tito, Vanleeuwen, Falony and Raes2017). The effect of drop-out on statistical measures is expected to be the same in clinical GMB trials. Despite drop-out being common in clinical studies, its effect on outcome measures involving microbial compositional data (e.g., beta diversity) is not currently well described in clinical GMB studies. However, it is expected that such measurements relying on consistent analysis from a wide array of samples will be biased if there is an inadequate sampling size.

The effect of bias comes into effect when there is an interpretation between samples, in that missing data prevents consistent interpretation of genetic data through a larger body of samples. For example, microbiome samples stratified by disease state versus control should be held to higher statistical power, similar to traditional clinical studies. Yet, the complexity of GMB genetic analysis often prevents large sample sizes from being a practical implementation due to costs unless utilising less-expensive protocols such as those involving qPCR to monitor microbial composition at high taxonomic levels (i.e., phyla) (Koliada et al., Reference Koliada, Moseiko, Romanenko, Piven, Lushchak, Kryzhanovska, Guryanov and Vaiserman2020). Some studies demonstrate shallow shotgun metagenomic sequencing as an alternative methodology for large, longitudinal GMB studies (Xu et al., Reference Xu, Chen, Pei, Guo, Li, Yang, Zhang, Yu, Li, Yang, Zhao and Wu2021). Nonetheless, making interpretations in GMB data between samples stratified by host conditions may need to be more consistent and accurate when samples are unavailable from a patient drop-out. Based on the literature of other areas in clinical research as discussed, it is again reasonable to assert that drop-out will influence outcome measures if authors make interpretations across hosts of varying condition states.

Due to few clinical studies analysing the effect of drop-out on GMB outcomes, it is reasonable to use a 20% patient drop-out ratio, as many clinical trials traditionally utilise. GMB studies that have a high patient dropout are considered high ROB. GMB studies that have low patient drop-out are considered low ROB.

Sequencing depth and sampling zeroes

GMB researchers should consider sequencing depth as a contributor to missing data and subsequent bias. It is established that low-sequencing depth (2000 single-end reads per sample) can adequately predict the same diversity patterns as high-depth sequencing (on the scale of millions of reads per sample) (Caporaso et al., Reference Caporaso, Lauber, Walters, Berg-Lyons, Lozupone, Turnbaugh, Fierer and Knight2011; Lundin et al., Reference Lundin, Severin, Logue, Ostman, Andersson and Lindström2012; Xiao et al., Reference Xiao, Yu, Li, Juneau and Yan2018). Experiments that quantify GMB outcome measures (like alpha and beta diversity) should utilise the same depth for all samples. Bias would be introduced if different sequencing depths are used for a set of samples. It should be noted, however, that false zeroes influence microbiome genetic data at both high and low depth. While true zeroes (or biological zeroes) represent true taxonomic absences, false zeroes (or sampling zeroes) represent a lack of sequencing depth to adequately detect certain microbial taxa. Notably, low sequencing depth, as is often the case of 16S rRNA sequencing, may not detect low abundance taxa or low taxa (subspecies) due to lower resolution. Though whole genome sequencing (WGS), such as shotgun metagenomic sequencing, utilises high sequencing depth to sequence entire genomes, sampling zeroes still persist (Pereira-Marques et al., Reference Pereira-Marques, Hout, Ferreira, Weber, Pinto-Ribeiro, van Doorn, Knetsch and Figueiredo2019).

At the time of writing, this issue of zero-inflation – or the excess of sampling zeroes at high and low depth – and the resulting bias in GMB genetic data is an active area of research. Interestingly, relatively few studies utilise any statistical modelling to correct for such missing data. Yet, various modelling techniques were recently developed to address zero-inflation (Ha et al., Reference Ha, Kim, Galloway-Peña, Do and Peterson2020; Zhang et al., Reference Zhang, Guo and Yi2020; Deek and Li, Reference Deek and Li2021). Similar to modelling techniques, imputation is a method traditionally used to address missing data in the form of patient drop out, but a promising imputation method is recently available to also deal with GMB sampling zeroes. Previous studies showed an increase in Pearson correlation from 0.59 (between 16S and WGS in non-corrected data) to 0.64 (between 16S and WGS in corrected data) (Jiang et al., Reference Jiang, Li and Li2021 ). There were also marked differences in mean and standard deviation of abundances per taxon between corrected and non-corrected data. This suggests greater homogeneity of samples across sequencing methods if imputation is utilised to correct data. However, as our article focuses on the role of bias in GMB research, we do not yet place best-practice recommendations for a particular method of missing data correction.

As of date, few GMB studies utilise statistical techniques to correct for sampling zeroes. Furthermore, common bioinformatics pipelines (such as QIIME2) do not incorporate such techniques into data-correction programs.

As such, the available literature suggests future GMB studies that do not consider sampling zeroes and lack a statistical technique for missing data correction may be considered high ROB. Studies that utilise missing data correction may be considered low ROB. These data correction methods, once more, include various modelling techniques or imputation.

Sample collection

Currently, there is no standard method for sample collection for GMB studies. While biopsy of the lower intestine provides a controlled sampling site and an accurate microbiota account, it is expensive, time-consuming, and unsuitable for healthy control groups. In contrast, the faecal collection is non-invasive and cost-effective (Tang et al., Reference Tang, Jin, Wang, Liu, Liu, Wang and Cao2020). Thus, it is a standard sampling method in both clinical and research applications. However, faecal sample collection introduces temporal inconsistency which can be a source of bias.

Faecal samples collected at different times of the day are at risk for inaccurate representation of the absolute abundance of gut microbiota (Caporaso et al., Reference Caporaso, Lauber, Walters, Berg-Lyons, Lozupone, Turnbaugh, Fierer and Knight2011). Specifically for mouse studies, the snapshots of the microbiota provided by the faecal samples are more accurate and consistent within treatment groups when collected in the morning due to the nocturnal feeding nature of mice (Jones et al., Reference Jones, Reinke, Ali, Palmer and Christophersen2021). For studies involving subjects with unpredictable and inconsistent bowel movements, samples should be preserved immediately after defecation as oxidation of the outer layer can alter the microbiota (Pepper and Rosenfeld, Reference Pepper and Rosenfeld2012). Specifically, Firmicutes and Bifidobacteria spp. are two known phylum that are unstable in the outer microenvironment when exposed to oxygen (Gorzelak et al., Reference Gorzelak, Gill, Tasnim, Ahmadi-Vand, Jay and Gibson2015). Therefore, to minimise the differential errors, the methods of measurement must be consistent between control and intervention groups.

Blinding

In a GMB study, the primary outcome is based on definitive and objective genetic sequencing. Therefore, assessor bias is typically negligible, and a low ROB is expected (Higgins et al., Reference Higgins, Thomas, Chandler, Cumpston, Li, Page and Welch2022).

Selection of reported results

Selective reporting of results can lead to biased interpretations of significance and or non-significance via particular selection of results from multiple outcome measures in estimating outcome effect. Bias in the selection of reported results can be difficult to detect without access to a protocol from which one can compare pre-specified intended outcomes of interest to the outcomes analysed in the published paper (Heneghan et al., Reference Heneghan2019). Often, results are selected for significance, omitted for non-significance, or omitted for adverse effect of intervention (Dwan et al., Reference Dwan, Gamble, Williamson and Kirkham2013; Hedin et al., Reference Hedin, Umberham, Detweiler, Kollmorgen and Vassar2016; Van der Steen et al., Reference Van der Steen, Ter Riet, van den Bogert and Bouter2019).