Description of the evidence and interventions

From the 12 studies identified by the process shown in Figure 8, six were RCTs^98–103 and six were observational studies.^104–109 They included 123,788 THRs and nine infection prevention strategies, as shown in the MTC network (Figure 9). The data from the included papers are shown in Table 6.

FIGURE 9

Mixed-treatment comparison network consisting of 12 studies, with nine infection prevention strategies. Note that the lines represent direct evidence comparisons, boxes represent infection control strategies involving multiple infection control measures (more...)

TABLE 6

Summary of evidence: nine infection prevention strategies across the MTC network

The quality of evidence summarised in Table 6 was variable, with quality score between 11 and 21 (see Tables 65 and 66 in Appendix 3 for the scoring tools used). Five of the six RCT studies^99–103 provided no information on random sequence generation, four^100–103 provided no information on blinding assessors and only one⁹⁸ reported prior calculation of the sample size. The statistical power for most RCTs was generally low. Only one⁹⁸ RCT reported primary analysis based on all randomised cases, whereas the rest did not report intention to treat. Of the six observational studies, three^105,108,109 identified and adjusted for confounding variables, one¹⁰⁶ reported that cases and control groups were comparable on diagnostic confounding factors and two^108,109 described and included in the analysis the outcomes of the patients who withdrew. Three studies^101,108,109 used objective measures to assess the outcomes and were adequately powered with large sample size ranging from 10,905 to 51,485.

Effectiveness outcomes

For every strategy in the connected network a relative effect was estimated against another infection prevention strategy using an OR of SSI. We chose ‘no systemic antibiotics, plain cement and conventional ventilation’ as the reference strategy (T1), as it was compared with the greatest number of other strategies.

Thirty-six relative effects involving nine infection prevention strategies were estimated in the MTC network using models that did and did not adjust for duration of follow-up (Table 7). The results from both models were almost identical, as were estimates of the model fit; therefore, the differences in follow-up duration had little effect on the effectiveness of the infection strategies. Therefore, we report the results of the model without adjustment for follow-up from now on. The 36 ORs for all pairwise comparisons are presented in the forest plot in Figure 10. The probability and median rank of a strategy being the most effective strategy is shown in Table 8. The models were fitted in a Bayesian framework using WinBUGS and code by Dias et al.¹¹⁰

TABLE 7

Odds ratios with 95% credible intervals of all infection prevention strategies

FIGURE 10

Forest plot of ORs of SSI for infection prevention strategies (random effects). Reproduced from Zheng H, Barnett AG, Merollini K, Sutton A, Cooper N, Berendt T, et al. Control strategies to prevent total hip replacement-related infections: a systematic (more...)

TABLE 8

Probability of each infection prevention strategy being the most effective

The five infection prevention strategies associated with a statistically significant reduction in THR-related SSI compared with the reference strategy (T1) were:

1. T6 with an OR of 0.13 [95% credible interval (CrI) 0.03 to 0.35]; systemic antibiotics, antibiotic-impregnated cement and conventional ventilation
2. T4 with an OR of 0.25 (95% CrI 0.06 to 0.66); systemic antibiotics, plain cement and laminar airflow
3. T3 with an OR of 0.26 (95% CrI 0.03 to 0.95); no systemic antibiotics, plain cement and laminar airflow
4. T7 with an OR of 0.27 (95% CrI 0.03 to 0.93); systemic antibiotics, antibiotic-impregnated cement and laminar airflow
5. T2 with an OR of 0.31 (95% CrI 0.12 to 0.65); systemic antibiotics, plain cement and conventional ventilation.

Statistically non-significant reductions in THR-related SSI compared with the reference strategy, T1, were:

• T5 with an OR of 0.38 (95% CrI 0.09 to 1.12); no systemic antibiotics, antibiotic-impregnated cement and conventional ventilation
• T8 with an OR of 0.52 (95% CrI 0.03 to 2.12); systemic antibiotics, antibiotic-impregnated cement, conventional ventilation and body exhaust suit
• T9 with an OR of 0.74 (95% CrI 0.05 to 2.69); systemic antibiotics, antibiotic-impregnated cement, laminar ventilation and body exhaust suit.

When T7 was compared with T6, the OR of SSI was 1.96 (95% CrI 0.52 to 5.37), suggesting that laminar airflow could increase infection risk. Similarly, when T8 was compared with T6, the OR was 3.72 (95% CrI 0.38 to 13.75), suggesting that body exhaust suits may also increase infection risk, at least where there is conventional ventilation. There was no high-quality evidence that antibiotic-impregnated cement without systemic antibiotics was more effective in reducing infection than plain cement and systemic antibiotics (T2 vs. T5; OR 1.28, 95% CrI 0.38 to 3.38). All comparisons and interpretations are summarised in Table 9.

TABLE 9

Control strategies to prevent THR-related infections: a systematic review and MTC

Model fit and evidence consistency

The model fit statistics indicate that the fit was less than adequate (see Table 7). This was confirmed by diagnostic plots that showed infection prevention strategies T2 and T5 of study 4¹⁰² and the T1 strategy of study 10¹⁰⁹ were outliers contributing to the inadequate model fit (Figure 11).

FIGURE 11

Leverage vs. deviance residual superimposed on curves y = –x² + c, where c = T1, T2, T3 and T4, representing the amount contributed to DIC. Reproduced from Zheng H, Barnett AG, Merollini K, Sutton (more...)

Curves of the quadratic function were plotted as they represented the lines of each contribution to DIC. Points lying outside the line c = 3, were identified as contributing to the inadequate model fit. The plot shows the first and second arms, strategies T2 and T5, of study 4¹⁰² are outliers contributing to the inadequate model fit.

After exclusion of both the first and second arms of study 4¹⁰² (T4 and T1, and T4 and T2, respectively), the model fitted the data well and the heterogeneity was significantly reduced, but the results were little changed (Figure 12).

FIGURE 12

Sensitivity analysis excluding the first and second arms of study 4 (T4 and T1, and T4 and T2, respectively). Leverage vs. deviance residual superimposed on curves: y = –x² + c, where c = T1, T2, (more...)

A further sensitivity analysis was undertaken that removed the first arm of study 10¹⁰⁹ (T1 and T10) (Figure 13). After excluding study 10¹⁰⁹ from the network, all the remaining data points lay below the quadratic curve with c = 3, suggesting that the contribution of the remaining data points to the DIC was unimportant, which in turn would improve the model fit.

FIGURE 13

Sensitivity analysis by further excluding the first arm of study 10 (T1 and T10). Leverage vs. deviance residual superimposed on curves: y = –x² + c, where c = T1, T2, T3 and T4, representing the (more...)

Following these sensitivity analyses, infection prevention strategy T6 remained dominant, with the highest probability (64%) and highest median rank of being the most effective strategy (Table 10).

TABLE 10

The probability of each infection prevention strategy being the best strategy and its median rank (sensitivity analyses)

The MTC results are shown in Table 11. The sensitivity analysis that excluded studies 4¹⁰² and 10¹⁰⁹ from the MTC network showed that model fit was improved; the DIC was reduced from 180.6 to 141.8. The posterior mean residual deviance was also reduced from 34.3 to 25.3. Heterogeneity measured in between-study standard deviation across the MTC network also reduced from 0.63 to 0.43.

TABLE 11

Odds ratios with 95% CrIs of all nine infection prevention strategies (sensitivity analysis)

The direct evidence from all conventional pairwise meta-analyses is presented in Table 12. There was broad agreement among the direct evidence from conventional pairwise meta-analyses, the direct and indirect evidence from node splitting and the evidence from the MTC model. Tests for inconsistency between direct and indirect evidence from node splitting suggested that there was no statistically significant evidence of inconsistency. The model fit statistics for the node splitting and the MTC models were similar, implying that there was no conflict between the direct and indirect evidence. It is worth noting that the 95% CrIs for some pairwise comparisons widened greatly following node splitting. This is explained by the node splitting reducing the evidence available to inform the variance.

TABLE 12

Evidence from the MTC of 10 studies (excluding studies 4 and 10 from the MTC network), direct evidence from pairwise meta-analysis, and direct and indirect evidence from node splitting (relative intervention effects are in log-OR)

A test of interaction between RCTs and observational studies was not statistically significant, suggesting that combining these study types was not inappropriate (Table 13).

TABLE 13

Metaregression on subgroup interaction between RCTs and observational studies

The results were little changed by excluding the RCT by Hill et al.⁹⁸ (Table 14) or by including the RCT by Lidwell et al.¹¹¹ (Table 15).

TABLE 14

Probability of each infection prevention strategy being the best, excluding Hill et al.

TABLE 15

Probability of each infection prevention strategy being the best, including Lidwell et al.

Strategy T6 (systemic antibiotics, antibiotic-impregnated cement and conventional ventilation) remained dominant in both these further analyses, with the highest probability of being a cost-effective decision (63% and 83%, respectively) and highest median rank of being the most effective strategy (see Tables 14 and 15).