Are we throwing away good data? Evaluation of chimera detection algorithms on long-read amplicons reveals high false-positive rates across algorithms

Simulated data

To test the chimera detection approaches on long reads, we first constructed a simulated PacBio dataset based on the reference 186 full-ITS (ITS1-5.8S-ITS2) sequences obtained from the EUKARYOME v.1.9.2 database Tedersooetal2024 related to the study of Jamy et al. (2022). The reference sequences, annotated at the species level, are listed in Table S1. The simulated dataset was generated with SimLoRD v1.0.4 (Stöcker, Köster & Rahmann, 2016), a package designed for simulating PacBio data with characteristic errors. For each of the 186 ITS sequences, we assigned a unique read count using the -n parameter in SimLoRD, calculated to achieve proportionate representation according to sequence properties per input. This included the creation of 1,000 to 9,000 full-ITS sequences per input sequence while maintaining sequence length variation typical of the original data. SimLoRD was executed with up to 30 passes to create enough variability and simulate sequencing complexity similar to a real-world setting (Table S2). SimLoRD outputs FASTQ-format reads with simulated per-base quality scores. The total pool of simulated dataset (424,010 sequences) was then subjected to quality filtering (VSEARCH --fastq_filter --fastq_qmax 93, --fastq_maxee 1), and followingly, ITSx v.1.1.3 (Bengtsson-Palme et al., 2013) was used to clip flanking primer binding sites by keeping only full-length ITS (--nhmmer TRUE, -E 1e-2, --complement TRUE, --only_full TRUE). These preprocessing steps eliminated low-quality, truncated reads and reduced the dataset to 44,470 full-length ITS sequences.

The latter dataset was supplemented with in-silico chimeras. The chimeric reads were generated using a custom Python script (Table S2) that creates chimeras by randomly selecting pairs of sequences, determining breakpoints based on ITS region-specific probabilities, and applying length constraints to ensure realistic sequence properties. In addition, it ensures that the chimeras do not dominate in abundance and length compared to the original (parent) sequences. A total of 2,484 chimeras were generated (Table S3). After adding chimeras to the simulated dataset, the total number of sequences was 46,954 (28,989 unique sequences and 6.2% chimeric reads). This approach provided a dataset where the origin of each sequence, whether parental or chimeric, is known.

Scoring chimera detection

To examine the precision and recall of uchime_denovo and chimeras_denovo, a series of runs with different parameter settings were executed with our simulated dataset. Here, sequences flagged as ‘borderline’ by uchime_denovo were considered chimeras. The removeBimeraDenovo function (of the DADA2 package) was tested only with default settings. Precision is the number of detected true chimeras (true positives) divided by the number of all chimeras returned by the chimera filtering process. This can be expressed as true positives/(true positives + false positives). Recall is the proportion of true positives detected correctly over the total number of values (number of sequences) classified as true positives and false negatives (true positives/true positives + false negatives). We used the F1 score, a measure that evaluates the performance of an algorithm by considering both precision and recall. ${\displaystyle F1=2\times \left(\frac{precision\times recall}{precision+recall}\right)}$

Various tested adjustable parameter settings are listed in Table 1. We tested a combination of 49 runs for uchime_denovo and 22 runs for chimeras_denovo.

Empirical dataset

We used a dataset from Jamy et al. (2022) (BioProject https://www.ebi.ac.uk/ena/browser/view/PRJEB45931) for testing chimera filtering algorithms on an empirical ITS amplicons dataset. The samples obtained in this investigation were derived from various environments, encompassing nine from marine (marine water), four from freshwater (three lake water, one pond water), one from freshwater sediment (lake sediment), and four from the terrestrial (two peat soil, two forest soil). The samples comprised 10,609,939 sequences amplified using the general eukaryotic primers 3NDf and 21R (Jamy et al., 2022). These primers targeted a segment of the 18S gene, the entire internal transcribed spacer (ITS), and a portion of the 28S gene. The amplicons were subsequently sequenced using PacBio’s Sequel II platform. No spike-in or positive control sequences were included in these samples. Before chimera filtration per-sample, all samples underwent primer trimming by cutadapt v4.4 (Martin, 2011), quality filtering with DADA2 v.1.32 (truncLen = 0, maxEE = 2, minQ = 3, rm.phix = FALSE), extraction of the full-length ITS region by ITSx (--nhmmer TRUE, -E 1e-2, --complement TRUE, --only_full TRUE), and dereplication by VSEARCH. Putative tag jumps were filtered out by UNCROSS2 (Edgar, 2018) within PipeCraft2 (Anslan et al., 2017) (options: f = 0.03, p = 1). We deployed 2,070,676 dereplicated ITS sequences for the chimera filtration processes (Fig. 1).

Detection of false-positive and false-negative chimeric reads

Reads filtered out during the chimera filtering step were subjected to BLASTn (Altschul et al., 1990) search (word size = 7; reward = 1; penalty = -1; gap opening cost = 1; gap extension cost = 2), using EUKARYOME v.1.9.2 as a reference database. When a query sequence had a high identity (≥99%) and high query coverage (≥99%) against the reference sequence, then the query read was considered a false-positive chimera. The remaining sequences in the chimeric set (after chimera filtering) were considered true chimeras. Reads that passed the chimera filtering process were also subjected to BLASTn search, where, in addition to EUKARYOME, the set of sequences from the corresponding sample (the sample from which the query originated) served as a reference database. To identify true chimeras, we first excluded self-hits, where the query and target sequence IDs were identical. For the remaining sequences, if a query had multiple alignments per best hit and the first high-scoring segment pair (HSP) had a query coverage below 85%, the sequence was classified as a true chimera. If such a chimera (sequence) was found within the non-chimeric output of the chimera filtering algorithms, it was flagged as a false negative chimera.

OTU-level comparisons between chimera filtering approaches

Output reads of each chimera filtration algorithm with default and adjusted settings (except removeBimeraDenovo) were clustered into operational taxonomic units (OTUs) using VSEARCH with a similarity threshold of 98%. Following this, chimeric and non-chimeric batches of sequences were subjected to BLASTn search against the EUKARYOME database to identify false-positive and false-negative chimeras. After recovering false positives and discarding false negatives from the chimera filtered dataset, the final OTUs set is referred to as “adjusted + FP-FN”. Clustering was done before chimera filtration to generate the “raw data OTUs”.

To account for differences in sequencing depth, we retained raw OTU count tables and converted OTU abundances to binary presence/absence (PA) data. Raw data OTUs were compared against those treated with various chimera removal methods, and default chimera filtering settings against the ones where false positives and false negatives were corrected. Furthermore, we performed all pairwise comparisons among chimera-filtering combinations (default vs. adjusted + FP-FN). For each table, sequence data were Hellinger-transformed using the decostand function from the vegan package v.2.6-8 (Oksanen et al., 2024) and Bray-Curtis dissimilarities computed using the vegdist function. After performing non-metric multidimensional scaling (NMDS), we used paired Procrustes tests to quantify discrepancies between ordinations. We assessed whether the impact of parameter tuning differed among chimera-filtering methods and sample types by fitting a linear mixed-effects model in lme4. The response variable was the Procrustes residual for each sample, and we treated both the filtering method (chimeras_denovo, removeBimeraDenovo, uchime_denovo) and the isolation source (forest soil, lake water, marine water, peat soil) as fixed factors. Sample identity was included as a random intercept to account for repeated measures across methods. We compared this full model to a reduced version that omitted the method × habitat interaction using AICc and a likelihood-ratio test (p < 0.001), confirming that the interaction term significantly improved model fit. Finally, we extracted marginal predictions and conducted pairwise contrasts among methods (holding habitat constant) to identify which algorithms showed the largest shifts in residuals when default parameters were replaced with adjusted + FP-FN settings. Visualizations were generated in ggplot2 v.3.5.1 (Wickham, 2016).

Kruskal–Wallis rank-sum test was utilized to assess whether different chimera removal methods had a significant effect on alpha diversity. Alpha diversity was assessed using Shannon’s index and the residuals from a linear model. Linear models were fitted with ‘chimera removal method’, ‘isolation source’ (e.g., soil, freshwater, marine), and their interaction as fixed effects. Residuals from these models were extracted and used as an adjusted alpha diversity measure. Statistical significance was defined at α = 0.05. All analyses were conducted in R v.4.4.2 (R Core Team, 2024) and RStudio v.2024.12.1.563 (R Core Team, 2024).

Characteristics of false-negative chimeras

To investigate why certain chimeric sequences escape detection (false negative chimeras), we examined their structural properties in the empirical dataset. We used Infernal v.1.1.5 (Nawrocki & Eddy, 2013) and the inferrnal R package v.0.99.8 (Furneaux, 2024) with the RF00002 covariance model from Rfam (https://rfam.org/family/RF00002) under default settings to identify the quantity and positions of the 5.8S rRNA gene within the ITS sequences. The 5.8S rRNA gene is a conserved component of the ITS region, and a genuine ITS sequence should contain exactly one 5.8S gene. However, if a sequence contains multiple 5.8S rRNA genes, it might be a chimera resulting from concatenating two or more ITS regions. We also examined the length distribution of the false negatives, as unusually long reads could further indicate concatenation artefacts.

Chimera detection in the simulated dataset

The default settings of uchime_denovo detected 1,975 (79.5%) out of 2,484 simulated chimeric reads with no false-positive detections (F1 score = 0.87; Fig. 2, Fig. S1). The algorithm yielded the highest chimera detection performance (F1 = 0.89; 1,991 true chimeras and two false positives; Table S4) when the parameters were changed to abskew 3 (default is 2) and minh 0.09 (default is 0.28), while all other settings remained default (Table S4; Fig. 2, Fig. S1).

The chimeras_denovo algorithm with default settings detected 1,199 (48.3%) true simulated chimeras, but it also filtered out many false-positive chimeric sequences (11,088 reads; F1 score = 0.15). The best performance of the chimeras_denovo algorithm was achieved by setting the abskew value to 4 (default is 1), the chimeras_length_min value to 30 (default is 10), and the chimeras_diff_pct value to 0.9 (default is 0). This detected 1,447 true chimeric reads and 143 false positives (F1 score = 0.71; Table S5). The undetected chimeras (i.e., false negatives) from both algorithms’ (uchime_denovo and chimeras_denovo) outputs did not demonstrate any other special characteristics besides a few being abnormally long (>1,000) or short (<100) sequences (Table S3).

The removeBimeraDenovo algorithm was tested only with the default settings, yielding 1,130 (45.5%) true chimeric and 12,764 false positive reads (F1 score = 0.14). It failed to filter out true chimeric sequences where one of the parts was in reverse complementary orientation (compared to the parent sequence) and sequences from closely related species.

Chimera detection in the empirical dataset

The uchime_denovo filtered out 12,114 reads (0.58%) as chimeric based on the default settings. Conversely, when the best setting (abskew 3, minh 0.09; as identified in the simulated dataset; adjusted settings) was applied, more reads were filtered out: 41,087 reads, comprising 2% of the dataset. The chimeras_denovo filtered out 498,437 reads (24.07%). Under adjusted settings (abskew 4, chimeras_length_min 30, chimeras_diff_pct 0.9), the proportion of discarded reads dropped notably to 99,673 sequences (3.85%). The removeBimeraDenovo algorithm with default settings filtered out 320,032 chimeric sequences (15.45%; Fig. 3). In total, only 151 reads were commonly flagged chimeric by all three methods with default settings; however, chimeras_denovo uniquely flagged more chimeras (270,207) than uchime_denovo and removeBimeraDenovo (Fig. 3). With adjusted settings for uchime_denovo and chimeras_denovo, 1,875 reads were commonly flagged as chimeras between these algorithms (Fig. 3).

Following chimera filtration, the chimeric sequences identified with default and adjusted settings were subjected to taxonomy annotation to identify false-positive chimeras. Using default settings, the false-positive rate in the putative chimeric reads set was 9.8% (1,052 reads) for uchime_denovo, 55.0% (276,346 reads) for chimeras_denovo, and 52.2% (167,102 reads) for removeBimeraDenovo (Fig. 4). Using adjusted settings, the false-positive rate increased to 19.9% (8,618 reads) for uchime_denovo, but it decreased to 38.6% (37,876 reads) for chimeras_denovo (Fig. 4). A comparative analysis of false positives across the three tools revealed that, using default settings, only 75 reads were shared among all methods (Fig. 5A). With adjusted settings (uchime_denovo & chimeras_denovo), the overlap increased to 90 reads, with removeBimeraDenovo (default settings) showing the largest count of unique false positives (Fig. 5C).

To determine the taxonomic composition of false-positive chimeras, we analyzed the phyla associated with each sample derived from the false-positive sequences identified by various chimera filtering methods. Using the uchime_denovo, sequences of Basidiomycota were most often misdetermined as chimeras across all habitats, with the highest relative rate in soil samples (Figs. S5A, S6A, Table S6). Ascomycota contributed strongly to false positives in freshwater and soil samples, while Glomeromycota were erroneously determined to be chimeric from soil samples (Figs. S5A, S6A, Table S6). In contrast, the chimeras_denovo method indicated that false-positive chimeras (Figs. S5B, S6B, Table S6) are mainly related to Dinoflagellata in freshwater samples, Arthropoda in soil samples, and Cercozoa in marine samples. The removeBimeraDenovo method misdetermined Basidiomycota and Ascomycota as chimeric in soil samples, but Annelida in marine samples. False-positive chimeras in freshwater samples were more evenly distributed among phyla (Figs. S5C, S6C).

The putatively nonchimeric outputs were evaluated for the presence of false-negative chimeras (chimeric sequences not identified during filtration; see ‘methods’). Under default settings, the rates of false-negative chimeras were consistent across various algorithms: 0.06% (1,314 from 2,058,562 reads) in uchime_denovo, 0.06% (1,275 from 1,572,239 reads) in chimeras_denovo, and 0.06% (1,297 from 1,750,644 reads) in removeBimeraDenovo (Fig. 4). With adjusted settings, these rates were slightly changed to 1,271 reads (total: 2,029,589, 0.06%) for uchime_denovo and 0.06% (1,291 from 1,971,003 reads) for chimeras_denovo. A comparison of shared false negatives showed that most sequences overlapped between methods (1,240 for default settings, and 1,238 for adjusted settings; Fig. 5). The empirical dataset’s false negative chimeric sequence length distribution exhibited a bimodal pattern with peaks at ca. 500 bp and 5,000 bp (Fig. S7). There was one group of sequences centered on the expected size of the ITS amplicon and another distinct group that consisted of extremely long reads, some longer than 3,000 bp. The results of Inferrnal for the 5.8S rRNA composition showed a relationship between sequence length and the number of 5.8S rRNA gene regions detected (Fig. S8). Most of these sequences were with two or more 5.8S hits. Specifically, 97.9% (1,218 of 1,244 reads) of false negatives from uchime_denovo and 95.5% from chimeras_denovo (1,218 of 1,275) with default settings. These figures remained high with adjusted settings, at 95.0% (1,207 of 1,271 reads) for uchime_denovo and 93.7% (1,210 of 1,291 reads). for chimeras_denovo. However, it was less pronounced in the removeBimeraDenovo, where 59.7% of false negatives (774 of 1,297 reads) had multiple 5.8s regions. These sequences had considerably higher median lengths (∼4,500 bp) than false negative sequences with no or one 5.8S hit (∼600 bp).

The impact of different chimera removal methods on OTU community structure was assessed using Procrustes analysis (Fig. S9). Procrustes correlations within each algorithm’s default versus adjusted + FP–FN (algorithm’s adjusted settings, as above, with recovering false positive and discarding false negative reads) pair remained very high (Procrustes R = 0.93–0.99, p = 0.001), with uchime_denovo demonstrating the strongest concordance (Procrustes R = 0.99; thus, lowest residuals, Fig. S10). With default settings, removeBimeraDenovo and chimeras_denovo filtered out high rates of false positives (52.2% and 55.0%, respectively), corresponding to the highest residuals (after Procrustes analysis) in the default versus adjusted + FP–FN comparison (Fig. S10). Similarly, all pairwise comparisons among default and adjusted + FP–FN chimera-filtering combinations yielded significantly high correlations (Procrustes R = 0.94 − 0.98, p = 0.001), but residuals were overall slightly lower between adjusted + FP–FN pairs (indicating reduced biases between algorithms; Fig. S10). The comparison of outputs from all chimera-filtering methods (both default and adjusted + FP–FN) with raw OTU tables (no chimera filtering performed) also showed significantly correlated sample positions in ordination space, but to a lesser extent than those observed among the chimera-filtered datasets (Procrustes R = 0.93–0.98, p = 0.001; Fig. S9A). The linear mixed-effects model was applied to the residuals of the default and adjusted + FP-FN comparison, including the interaction between the chimera filtering method and the isolation source (Fig. S11). Both the chimera removal method (F = 32.16, p < 0.001) and the isolation source (F = 11.54, p < 0.001) were significant factors. Importantly, the interaction between the method and the source was also highly significant (F = 11.90, p < 0.001). The chimeras_denovo had the largest residuals, particularly in forest soil (Estimate = 0.55) and lake water (Estimate = 0.58), whereas uchime_denovo was consistent, and residuals remained close to zero. Moreover, average pairwise comparisons confirmed that uchime_denovo had significantly lower residuals than both chimeras_denovo (Estimate = −0.24, p < 0.001) and removeBimeraDenovo (Estimate = −0.15, p < 0.001).

Kruskal–Wallis rank-sum tests revealed no significant effect of the chimera removal method on overall OTU richness or the Shannon diversity index (Figs. S12, S13). Across all three chimera filtering algorithms, correcting false positives mainly affected the sequence abundance of dominant OTUs. Nevertheless, this process recovered hundreds of unique OTUs, but most could only be assigned to higher taxonomic levels (e.g., kingdom or phylum). Specifically, for chimeras_denovo adjusted + FP-FN, 158 OTUs were gained, yielding three additional genera (from kingdoms Rhizaria, Haptista, and Metazoa; the remaining 155 OTUs classified to the kingdom/phylum level; Table S7). For removeBimeraDenovo + FP-FN, only 13 new OTUs were gained, adding a new metazoan genus. The largest OTU gain arose within the uchime_denovo adjusted + FP-FN dataset, which hosted 346 unique OTUs, and added four new fungal species and one genus (Table S7).

Performance insights from parameter optimization

The default settings of uchime_denovo demonstrated relatively high precision in the simulated dataset, where adjustments resulted in a modest gain (Fig. 2). On the other hand, the efficacy of chimeras_denovo was considerably influenced by calibration. Its default settings generated a remarkably elevated incidence of false positives in the simulated dataset, which resulted in low precision performance (Fig. 2). Thus, the parameter adjustments of chimeras_denovo notably increased the performance. Here, we applied removeBimeraDenovo only with default settings, and its performance was similar to the outcome of the chimeras_denovo with default settings (Figs. 2, 4). This indicates that the default parameters of chimeras_denovo and removeBimeraDenovo may be too aggressive by incorrectly classifying natural sequence variation as chimeric ones.

A recent benchmarking study has shown that filtering PacBio HiFi reads with adjusted parameters of uchime3_denovo VSEARCH achieves high accuracy in resolving ASVs with a minimal number of false positive chimera detections (Overgaard et al., 2024). Similarly, our findings suggest parameter adjustments can improve precision while maintaining recall. However, our study also highlights that parameter optimization is not always straightforward. Switching from default to adjusted settings frequently improved performance by reducing false negatives, but mostly at the expense of increasing the rates of false positives. We also found that performance depends on the specific dataset, as optimal adjustments on our simulated dataset did not directly translate to the same efficiency on the empirical data. Similar context-dependent variability has been reported previously, where Edgar et al. (2011) demonstrated that UCHIME’s filtering efficiency can decline when applied to highly complex environmental samples, particularly where sequence similarity among parental taxa and chimera breakpoints complicates detection. Additionally, in the empirical dataset, the differences between corresponding sample points as measured in an OTU community ordination space (Procrustes residuals) revealed the significant effect of the chimera filtering method and isolation source. The method-specific nature of chimera detection was also demonstrated on the limited overlap of identified chimeric sequences among algorithms (Fig. 3).

Secondary validation

Annotating the discarded sequences by chimera filtering algorithms revealed that a substantial proportion of those were falsely labeled as chimeras (Fig. 4). The same was found from ITS2-based mock community analyses, where high rates of falsely identified chimeras were identified using UCHIME (Aas, Davey & Kauserud, 2017). In our study, surprisingly, an algorithm originally optimized for long-read data, chimeras_denovo, detected more false-positive chimeras than uchime_denovo. Similarly, the removeBimeraDenovo yielded high rates of false-positive chimeras, which could be ‘rescued’. However, the correction for false positives across all three algorithms resulted only in a slight gain of new OTUs that would have been discarded without secondary validation of false positives. Correspondingly, the alpha diversity metrics were not significantly different between chimera filtered datasets that either incorporated the rescue of false positives or not (Fig. S12). Similarly, the Procrustes analyses between all pairs of chimera-filtered datasets demonstrated high and significant correlations (Procrustes R = 0.94−0.98, p = 0.001), indicating no significant impact of false-positive chimeras’ exclusion on broad-scale community analyses. Likewise, analyzing the performance of UCHIME upon fungal ITS2 sequences, Aas, Davey & Kauserud (2017) concluded satisfactory results with the default chimera filtering approach.

The proportion of identified false-negative chimeras in the filtered datasets was notably smaller compared with false positives. Many false negatives were unusually long and, in several cases, contained multiple hits to the 5.8S rRNA region. This suggests that some chimeras in our dataset were not generated during PCR amplification but likely emerged during PacBio library preparation. The formation of long chimeric constructs during SMRTbell adaptor ligation can evade standard chimera detection algorithms (Fichot & Norman, 2013; Griffith et al., 2018). Although sequences outside the expected amplicon length can be filtered out during, for example, the quality filtering step, our analysis did not apply a maximum length threshold (because of the highly variable length of ITS region across eukaryotes; Yang et al., 2018). However, the relative abundance of these artifacts was low, suggesting that they hinder the community-level analyses even less than false positives. Nevertheless, reporting those as valid OTUs may create undesirable biases in precise richness analyses and generate erroneous metabarcoding feature entries when those are deposited, for example, in PlutoF (or similar) biodiversity data management platforms (Abarenkov et al., 2010).

The utility of secondary validation aligns with findings from studies that emphasize the value of hybrid approaches in improving long-read data analyses (Christel et al., 2023; Lu et al., 2023; Rué et al., 2023). Therefore, few metabarcoding pipelines, such as PipeCraft2 and BIOCOM-PIPE (Djemiel et al., 2020), have implemented false-positive chimera rescue methods based on the secondary validation against a reference database. However, the reliance on reference-based validation is hampered for taxa without reference sequences. Another strategy implemented in the Natrix (Welzel et al., 2020) and NextITS pipeline (Mikryukov, Anslan & Tedersoo, 2025), for false-positive chimeras, involve retrieving sequences initially considered as chimeras but are present in multiple samples because the formation of identical chimeras in multiple independent PCR reactions is highly unlikely. This strategy, however, requires careful filtering of tag-switching errors to avoid rescuing true chimeras that happen to leak to multiple samples. With the rising awareness that error-free chimera filtering is not achievable (Edgar, 2016), we can expect various secondary validation strategies to be incorporated into metabarcoding pipelines in the future.

In addition to secondary validation strategies and/or adjusting chimera filtering algorithm parameters, using upstream solutions in library preparation can be beneficial for the chimera burden. For instance, implementing Unique Molecular Identifiers (UMIs) in the amplicon sequencing process can improve chimera removal by filtering out non-matching UMI amplicons, regardless of the long-read sequencing technology used (Karst et al., 2021). Furthermore, the use of dual UMIs eliminates chimeras during the data processing step, making the de novo filtering discussed here less critical for both PacBio and Oxford Nanopore Technologies (ONT) when per-UMI coverage is adequate (Overgaard et al., 2024).

Implications of fine-tuning the chimera filtering process

We found that the chimera filtered vs. non-filtered datasets demonstrated significant and high correlations in an ordination space (Procrustes analyses on NMDS axes). Therefore, although the broad-scale community structure in an ordination space remains largely preserved, chimera filtering is advisable (Tedersoo et al., 2022). Further fine-tuning of the algorithm’s parameters, including correcting for false positives and false negatives, however, seems to have little impact on the final overall data structure. Importantly, the degree of correspondence between OTU datasets processed with default versus adjusted (+FP-FN) chimera filtering settings varied depending on the DNA isolation source (i.e., sampling substrate; Fig. S9). As noted above, chimera filtering may be more challenging in high-complexity environmental samples; therefore, suggesting higher importance of secondary validation in, for example, forest soil datasets compared with marine water datasets (Figs. S9, S11). While our study focused on PacBio HiFi reads, the difficulties of de novo chimera detection may likely extend to reads generated with other sequencing technologies with greater challenges expected in datasets with more complex communities. Taken together, these results indicate that while refining the chimera filtering process does not radically alter the overall structure of the data, secondary validation steps help to balance the methodological biases.