We undertook analyses in the UK Biobank European population as illustration [11]. In UK Biobank, each individual was asked to perform up to three blows on a Vitalograph spirometer (Vitalograph Pneumotrac 6800). These blows were used to derive measures of lung function, including FEV₁, FVC and FEV₁/FVC, from a spirogram (volume-time curve) recorded by the Vitalograph software. A blow received an automated error code from the spirometry software if: (1) there was hesitation (excessive extrapolated volume [12] at the start of the blow, coded “START”); (2) the time to peak flow was excessive (“EXPFLOW”); (3) a cough was detected during the manoeuvre (“COUGH”); (4) there was not an adequate plateau at the end of the blow (“END”) (see Table 1 for additional detail). It was also possible for the spirometer operator to explicitly reject the blow (“REJECT”). In previous GWAS of lung function, blows were deemed unacceptable and excluded from analysis if they had any of the above error codes [4].

Additionally, as in previous GWAS [3, 4], we checked each blow for inappropriate negative values which indicate a problem with the blow (for any of: forced expiratory flow at 25% of FVC; forced expiratory flow at 75% of FVC; average forced expiratory flow between 25% and 75% of the FVC; peak flow; extrapolated volume; volume at forced expiratory time) (coded “NEGATIVE”), the start of the blow underwent a further check for hesitation (“START2”), and consistency (<5% difference) was checked between the FEV₁ and FVC output from the spirometer and that rederived from the spirogram (“CONSISTENCY”). In this study, we identified blows that would previously have failed QC due to any of the above error codes or additional QC steps, and labelled these with corresponding “acceptability flags” (Table 1).

We calculated the GRSs for European ancestry individuals in UK Biobank using European-specific weights trained from Shrine et al. [3] (Supplementary Methods). In the UK Biobank European population, we then conducted iterative testing of the GRS association within groups of spirometry blows stratified by acceptability flags, to identify and rank the flags most likely to cause failure of spirometry blow quality (Figure 1). All the GRS and lung function traits association tests were adjusted for age, age², height, smoking status, and top 10 genetic principal components. For each lung function trait, we first stratified all the blows by acceptability flag (Figure 1, step I, Supplementary Methods). Where lung function measures within a stratum showed no significant association with the GRS (P > 0.05), we considered that the corresponding acceptability flag to indicate unacceptable blow quality and thus a reason for exclusion. For the remaining blows, we applied an iterative selection process to identify the acceptability flags that were most likely to cause failure of spirometry blow quality as shown in Figure 1 (Step II). Based on the outcome of this, we ranked the remaining acceptability flags according to their impact on the spirometer blows in our new approach (see Results for details).

Based on the grouping of blows above, we then aimed to identify which acceptability and repeatability criteria would maximise the association of the lung function measures with GRS as measured by the level of statistical significance (Z-score for the association) (Figure 1, Step III). We varied our acceptability criteria by including blows with acceptability flags in order of the ranking generated in the previous step, in addition to blows previously considered acceptable. We also varied our repeatability threshold from 150ml – the standard recommended in strict clinical practice by ATS/ERS [8] - up to 400ml in 50ml increments. Repeatability was based on the difference from any other blow, even if that blow was not accepted. Using these different criteria, we tested the association with the GRS. To ensure that effect size estimates retained acceptable precision while maximising statistical significance, we then tested genetic associations with the sentinel SNPs – that is, the SNPs showing the strongest statistical associations with previous lung function traits GWAS in genomic regions [3]. Using the different acceptability criteria and repeatability thresholds described above, we examined how these QC criteria influenced the estimated effect sizes and P values for the sentinels of lung function signals. All lung function traits were untransformed, adjusting for age, age², height, smoking status, and relatedness (mixed models in BOLT-LMM [13]).

To illustrate the gain in power using our newly defined QC criteria in epidemiological studies, we conducted a GWAS in the UK Biobank European population as an example to compare GWAS findings with those found using the previous QC criteria. Credibility of newly identified signals was assessed using a Bayesian framework previously described by Okbay et al [14]and Turley et al. [15] (Supplementary methods).

Using the previous, more stringent approach to spirometry QC, we identified 713,885 spirometer blows as “accepted” for 354,746 European individuals. The best measures of FEV₁/FVC derived from “accepted” blows for these individuals were strongly associated with the GRS (β_{SD_change_of_GRS} =0.246, 95% CI [0.243, 0.249], P < 1.00E-300) (Figure 2a).

From the association of GRS with the FEV₁/FVC derived from blows in each of our acceptability flag strata, blows with acceptability flags “CONSISTENCY”, “NEGATIVE” or “REJECT” were not associated with the GRS (Figure 2a), indicating that these acceptability flags represent unacceptable spirometry blow quality (β_{SD_change_of_GRS} = –0.022, 95% CI [–0.049, 0.006], p = 0.1292, Figure 2c, Group 3). To rank the remaining acceptability flags that cause failure of the spirometry blow quality in descending order of severity, we conducted an iterative selection process (Figure 2b). In the first round of selection (round 1), we identified excessive time to peak flow (“EXPFLOW”) as the next most likely cause of failure of spirometry quality. This was based on the observation that additionally removing blows with “EXPFLOW” flag led to an increase in the magnitude of the effect size estimate in the association results in all the remaining strata. Similarly, we identified an inadequate terminal plateau (“END”, round 2) and cough (“COUGH”, round 3) as the next most likely to cause failure of spirometry quality in the subsequent rounds of selection.

Collectively, blows from groups “EXPFLOW”, “END” and “COUGH” were associated with GRS but with a smaller effect size than previously accepted blows (β_{SD_change_of_GRS} =0.149, 95% CI [0.145, 0.153], p < 1.00E-300, Figure 2c, Group 2), after removing blows with the flags excluded in previous steps. For the remaining flags relating to hesitation (“START” and “START2”), the magnitude of the effect size of GRS association in the corresponding strata (β_{SD_change_of_GRS}=0.262, 95% CI [0.240, 0.284], P=1.14E-118, Figure 2c, Group 1) was similar to that from the acceptable blows (β_{SD_change_of_GRS}=0.246, 95% CI [0.243, 0.249], P < 1.00E-300, Figure 2c, Accepted blows), after removing blows with all the other acceptability flags. For FEV₁ and FVC, we observed similar results to those for FEV₁/FVC above in ranking the spirometer blow quality but with the “COUGH” flag showing a similar effect size to accepted blows (Supplementary Figure 3 and Supplementary Figure 4). For FVC, “END” had a larger impact than “EXPFLOW”. Since the results suggest acceptability flags have greater impact on the measurements of FEV1/FVC than FEV1 and FVC, we based our subsequent analyses on FEV₁/FVC.

We investigated the trade-off between sample size and measurement error by re-evaluating various inclusion criteria for spirometry QC. To do this, we included blows with varying acceptability flags with their inclusion ordered according to the findings above, and applied different repeatability thresholds to assess their impact on the association between GRS and FEV₁/FVC. We found that including blows previously excluded for cough (COUGH), hesitation (START/START2), excessive time to peak flow (EXPFLOW) or lack of terminal plateau (END) in addition to accepted blows and applying a repeatability threshold of 350ml reached the maximum statistical significance in the association between GRS and FEV₁/FVC. However, as expected, the effect size in the association results attenuated toward zero as QC criteria were successively relaxed (Table 2).

To balance accuracy of effect size estimation versus maximising statistical significance (z-score) for GWAS, we examined the changes in the effect sizes and P values of the sentinels of FEV₁/FVC signals identified by Shrine et al. [3]. We found that additionally including blows previously excluded for cough (COUGH), hesitation (START/START2), excessive time to peak flow (EXPFLOW) or lack of terminal plateau (END), and applying a repeatability threshold of 250 ml optimised the signal-to-noise ratio in genetic association testing in UK Biobank (Figure 3). Based on this finding, we proposed a new spirometry QC strategy for epidemiological studies in UK Biobank, which retained 29% more participants using strictest ATS/ERS guidelines and increased the sample size from 275,084 to 356,053 individuals.

Using the newly defined spirometry QC criteria, the UK Biobank sample size increased from 320,591 in the most recent GWAS of lung function to 356,053, an 11% gain in sample size, and the mean χ² statistic (i.e. squared z scores for SNP and FEV₁/FVC association) from GWAS increased from 1.27 to 1.29, leading to 15 additional sentinel SNPs associated with FEV₁/FVC (selected 2 Mb regions centred on the most significant variant for all regions containing a variant with P < 5 × 10^-9), additional compared with analysis of UK Biobank alone using previous QC criteria [3]) (Supplementary Table 2). Of these, 7 were not identified in the largest GWAS of lung function to date. Examining the nearest genes to the sentinel SNPs, we found two novel genes in addition to 13 genes reported either in the most recent GWAS [3] or the EMBL-EBI GWAS Catalogue [17]. One of the novel genes, FPR3, encodes the formyl peptide receptor 3, a paralog of formyl peptide receptor 1. The functional role of FPR3 is not fully understood. However, it is expressed in a range of immune cells, including macrophages and eosinophils, but not neutrophils, so has been hypothesised to play a role in allergic disease [18], and has also been associated with asthma and white blood cell counts in GWAS. We also found that enrichment of a previously implicated pathway [3], ESC pluripotency pathway, was strengthened by the newly identified gene WNT16.

For the newly identified hits, we followed procedures previously described by Okbay et al. [14] to calculate the posterior probability of true association, which exceeded 99% for all 15 additional loci for any assumption about the prior probability of non-null SNPs in the range 1% to 99%. To test the sentinel SNPs for replication, we used meta-analysis results from 42 European ancestry cohorts (excluding participants from UK Biobank) of 249,114 individuals generated by Shrine et al [3]. Due to the limited statistical power for replication of genome-wide significant association with a smaller sample size, we applied methods to assess the replication of the effect size of sentinel SNPs. For the set of 13 newly identified sentinel SNPs for FEV₁/FVC available in the meta-analysis results, we regressed the effect sizes in UK Biobank on the effect sizes in the meta-analysis results with intercept constrained to be zero, after correcting the UK Biobank effect size estimates for winner’s curse bias using the method described in Turley et al [15]. The regression slope was 0.75 (standard error = 0.1376), being statistically significantly greater than zero (one-sided P=1.474 × 10^-4) but not statistically distinguishable from one (one-sided P = 0.0943), suggesting that the newly identified sentinel SNPs were replicated in independent datasets.