Results of the search

The first publication of this review contained 20 trials (Demicheli 1999). The second publication added five more trials (Demicheli 2004). The third publication included 48 trials in total (Jefferson 2007). The fourth published update (Jefferson 2010) included two new trials (aa Beran 2009a; aa Beran 2009b) and excluded three new trials (Belongia 2009; Chou 2007; Khazeni 2009). In this 2013 update, 41 new study reports have been included and 63 new trials have been excluded.

Some of the included studies had more than two arms, comparing different vaccines, routes of administration, schedules or dosages, or reported data from different settings and epidemic seasons. We split these studies into sub‐studies (data sets). For the remainder of this review, the term 'study report' refers to the original study report, while the word 'data set' refers to the sub‐study; these sub‐studies could refer either to different study arms, to different influenza seasons or to different study designs. Risk of bias can be independently assessed for each sub‐study (or data set) study design.

More information about the division of study reports into data sets is given in the Characteristics of included studies table. In this 2013 update, 90 studies (116 data sets) are now included in the review (Figure 1).

Figure 1

---

---

Study flow diagram

Included studies

We have coded each trial on the basis of study design and the type of data contributed to the review as follows. The letter preceding the study represents the study design: (a) denotes RCTs, (b) denotes case‐control studies and (c) denotes cohort studies. The second letter indicates the contribution to the evidence in the data set: (a) efficacy/effectiveness or (b) harms. So, for example, a case‐control study contributing safety or harms data is coded as (bb) and a trial contributing efficacy/effectiveness data is coded as (aa). A (p) code has been added to refer to the studies on vaccination during pregnancy.

Excluded studies

We excluded 155 studies (see Characteristics of excluded studies table).

Allocation

In the included trials allocation concealment was adequate (low risk of bias) in 18 studies (26.1%), inadequate (high risk of bias) in six (8.7%) and unclear (unclear risk of bias) in 45 (65.2%).

Blinding

We judged blinding as low risk of bias in 16 RCTs (23.2%), as high risk of bias in two studies (2.9%) and as unclear in 51 studies (73.9%).

Incomplete outcome data

The majority of the included RCTs/CCTs did not report sufficient information about loss to follow‐up (63 studies; 91.3%).

Selective reporting

The assessment of selective reporting bias presents several difficulties and would require review of the original study protocols for the included studies, which are mainly unavailable.

Other potential sources of bias

Few studies reported information on influenza circulation in the surrounding community, making interpretation of the results and assessment of their generalisability difficult.

It is now known that industry funding of influenza vaccine studies determines publication in high‐prestige journals and higher citation rates than other types of funding. In addition, industry funding is associated with optimistic conclusions, but the quality of the majority of influenza vaccine studies is low, irrespective of funding (see Table 2). A previously cited review showed a complex web of interrelationships between these variables (Jefferson 2009a), but how this impacts on policy‐making is not known.

Inactivated parenteral vaccines (Analysis 01)

The overall effectiveness of parenteral inactivated parenteral vaccine against influenza‐like illness (ILI) is 16% (95% confidence interval (CI) 5% to 25%), with a corresponding number needed to vaccinate (NNV) of 40 (95% CI 26 to 128). Heterogeneity amongst the studies in this comparison is relatively low (I² statistic = 26%) and a sensitivity analysis made by comparing estimates obtained using the random‐effects model versus the fixed‐effect model does not change the conclusion. The CI of the NNV becomes narrower by applying the fixed‐effect model (NNV 38, 95% CI 29 to 49).

Inactivated parenteral vaccines are 16% effective (95% CI 9% to 23%) in preventing ILI symptoms when strains contained in the vaccine antigenically match those circulating (Analysis 1.1.1). The estimated NNV for this comparison is 17 (95% CI 12 to 29). On the other hand, inactivated vaccines are not significantly protective against ILI when the degree of matching between the vaccine and circulating influenza strains is absent or unknown (risk ratio (RR) 0.90, 95% CI 0.69 to 1.18, Analysis 1.1.2). In the subgroup Analysis 1.1.2 heterogeneity is particularly high (I² statistic = 82%) and estimates using the fixed‐effect model show statistical significance: Vaccine Effectiveness 18% (95% CI 10% to 25%) and NNV 59 (95% CI 43 to 106).

The overall efficacy of inactivated vaccines in preventing confirmed influenza (Analysis 1.2) is 60% (95% CI 53% to 66%) with a NNV of 71 (95% CI 64 to 80). When the vaccine content matches the circulating strain, the efficacy is 62% (95% CI 52% to 69%) and the NNV is 58 (95% CI 52 to 69). The results are very similar when matching is absent or unknown (Vaccine Efficacy 55%, 95% CI 41% to 66% and NNV 60, 95% CI 50 to 80). Since heterogeneity was very low (I² statistic = 17% for Analysis 1.2.1; I² statistic = 14% for Analysis 1.2.2 and I² statistic = 11% overall), there were no differences when comparing the estimates obtained by using a fixed‐effect model with those from a random‐effects model.

Looking at the NNV in Analysis 1.1 and Analysis 1.2, it seems that effectiveness against ILI is higher than efficacy against laboratory‐confirmed influenza (NNV‐ILI 40; NNV‐influenza 71). These paradoxical results, showing an apparently higher aspecific effectiveness and a lower specific efficacy, are mainly due to the fact that ILI and confirmed influenza have a very different incidence among the study population. We note that 15.6% of unvaccinated participants versus 9.9% of vaccinated participants developed ILI symptoms, whilst the corresponding figures for participants who developed laboratory‐confirmed influenza are 2.4% and 1.1% for unvaccinated and vaccinated people, respectively.

Based on the results from a single study (aa Bridges 2000b), physician visits appear 42% less frequent (95% CI 9% to 63%) in participants immunised with vaccines prepared with strains matching circulating viruses (Analysis 1.3.1), whereas there are no significant results when the degree of matching is unknown or absent (RR 1.28, 95% CI 0.90 to 1.83; Analysis 1.3.2). The overall effect is also not significant (RR 0.87, 95% CI 0.40 to 1.89) (Analysis 1.3). Even though the two data sets of aa Bridges 2000b showed very high heterogeneity (I² statistic = 87%), no difference arose when comparing the results from the fixed‐effect with the random‐effects model analysis.

A similar conflicting result is observed when analysing the effect of inactivated vaccine administration on days of illness (Analysis 1.4), when the estimate (mean difference (MD)) obtained in good match conditions was compared with that where there was an unknown or absent degree of matching. As a consequence of the high overall heterogeneity (I² statistic = 87%), the result obtained from the fixed‐effect model analysis (MD ‐0.31, 95% CI ‐0.54 to ‐0.07) differs substantially from that resulting from the application of a random‐effects model (MD ‐0.21, 95% CI ‐0.98 to 0.56).

There seems to be no effect on the time an antibiotic or drug was prescribed (Analysis 1.5; Analysis 1.6).

Four trials evaluated time off work, estimating that vaccination saves around 0.04 working days on average. This result is affected by high levels of heterogeneity (I² statistic = 82%) and changes depending on whether a fixed‐effect (MD ‐0.04, 95% CI ‐0.06 to ‐0.01) or random‐effects model (MD ‐0.04, 95% CI ‐0.14 to 0.06) is used.

The effect on hospitalisation (Analysis 1.8) was evaluated in two trials (aa Bridges 2000a; aa Leibovitz 1971), but it was not statistically significant. No evidence was found for cases of pneumonia.

Live aerosol vaccines (Analysis 02)

Live aerosol vaccines have an overall effectiveness of 10% (95% CI 4% to 16%; NNV 46, 95% CI 29 to 115) and content and matching appear not to affect their performance significantly (Analysis 2.1). Overall efficacy (Analysis 2.2) is 53% (95% CI 38% to 65%) and the NNV is 39 (95% CI 32 to 54). Again, neither content nor matching appear to affect their performance significantly.

No evidence is available on complications (e.g. bronchitis, otitis, pneumonia).

The effectiveness of the aerosol vaccines against ILI (with no clear definition) is significant only for vaccines with absent or unknown matching (37%, 95% CI 20% to 51%) and the NNV is 69 (95% CI 23 to 46) (Analysis 2.3).

The conclusions of this comparison were unaffected by analysis using either the fixed‐effect or random‐effects models.

Inactivated aerosol vaccines (Analysis 03)

No RCTs assessing the effectiveness of inactivated aerosol vaccines in preventing ILI could be included; the only available evidence comes from studies carried out during the 1968 to 1969 pandemic (Analyses 12 to 16).

The efficacy of inactivated aerosol vaccine in preventing laboratory‐confirmed influenza (Analysis 3.1.1) is assessed in one RCT (aa Langley 2011), whose results do not show a statistically significant protective effect (RR 0.38, 95% CI 0.14 to 1.02).

Inactivated parenteral vaccines ‐ cohort studies (Analysis 04)

In this analysis, we have considered the effects of vaccine administration in pregnant women and their newborns.

Based on unadjusted data from a cohort study (high risk of bias), 2009/2010 H1N1 monovalent pandemic vaccines (Analysis 4.1.1) provide a significant protective effect against ILI in pregnant women (Vaccine Effectiveness 89%, 95% CI 79% to 94%; NNV 54, 95% CI 51 to 61). Seasonal inactivated vaccine is not effective against ILI (RR 0.54, 95% CI 0.22 to 1.32; Analysis 4.1.2). Sensitivity analysis performed using the fixed‐effect model showed statistical significance, even for a modest, protective effect (RR 0.76, 95% CI 0.65 to 0.89; NNV 92, 95% CI 63 to 201; VE 24%, 95% CI 11% to 35%).

The effectiveness of vaccination with seasonal inactivated vaccine during pregnancy for preventing ILI in newborns is not statistically significant, as it results from two cohort studies using either hazard ratio (HR) or RR adjusted estimates (Analysis 4.2.1 and Analysis 4.3.1, respectively). Efficacy against confirmed influenza (Analysis 4.3.2) is really modest but has statistical significance (adjusted RR 0.59, 95% CI 0.37 to 0.94; NNV 27, 95% CI 18 to 185; VE 41%, 95% CI 6% to 63%).

It seems that vaccination with the 2009/2010 H1N1 monovalent pandemic vaccine during pregnancy is not associated with a higher risk of abortion (Analysis 4.4.1 and Analysis 4.4.2), congenital malformation (Analysis 4.4.3) or neonatal death (Analysis 4.4.5).

Cases of neonatal death and abortion have been observed less frequently among women immunised with seasonal influenza vaccine (Analysis 4.5.1 and Analysis 4.5.4, both unadjusted estimates).

The results of pcb Deinard 1981 are based on the follow‐up results of 189 pregnant women immunised with monovalent pandemic A/New Jersey/8/76 (either in split or whole virus formulation) and 517 pregnant women who did not receive vaccination. The time of observation was extended up to the first eight weeks of life of the newborns. No statistically different incidence of maternal pregnancy outcomes and infant deaths was observed between vaccinated and unvaccinated groups. Statistical analysis (Chi² test) shows no relation between immunisation history and presence of anomalities at the 8th week of life. This cohort study has not been included in the analysis as the vaccine studied is no longer in use.

Inactivated parenteral vaccines ‐ case‐control studies (Analysis 05)

This analysis only includes studies assessing the effect of vaccination against influenza during pregnancy. The incidence of ILI in pregnant women who were immunised with inactivated seasonal vaccine during pregnancy was not statistically different when compared with that observed among unvaccinated pregnant women (Analysis 5.1.1). However, in sensitivity analysis using the fixed‐effect model, the results of the analysis become statistically significant. In conclusion, the results of this comparison were affected by the model used to perform the analysis.

One further case‐control study did not find a statistically significant association between exposure to seasonal inactivated vaccine in pregnancy and abortion cases (Analysis 5.2.1).

One retrospective cohort study tried to assess the effect of live attenuated vaccine during pregnancy, based on data from a health insurance database during six subsequent influenza seasons (pcb Toback 2012). A total of 834,999 pregnant women were identified, out of whom 138 received live attenuated vaccine at any time during pregnancy. Claims for hospitalisation or visits to the emergency department within 42 days after immunisation were searched for but all observed events were considered to be related to a normal physiological pregnancy and not to immunisation. The system used (claim data) would be unable to detect birth outcomes.

Serious adverse events ‐ Guillain‐Barré syndrome ‐ cohort studies (Analysis 06)

The possible association between exposure to seasonal inactivated vaccine in healthy adults and Guillain‐Barré syndrome onset within six weeks following immunisation was investigated by two cohort studies performed during two subsequent epidemic seasons. No significant association was found Analysis 6.1.1). Administration of seasonal inactivated vaccine during pregnancy was not associated with Guillain‐Barré syndrome onset within six weeks from immunisation (Analysis 6.1.2).

The cohort of cb Shonberger 1979 was the first study that compared Guillain‐Barré syndrome cases by vaccination status and the national incidence in vaccinated and unvaccinated national cohorts after the suspension of the National Influenza Immunisation Program in the winter of 1976 to 1977. At that time the monovalent inactivated swine vaccine A/New Jersey/8/76 had been administered. The attributable risk from vaccination was just below one case of Guillain‐Barré syndrome in every 100,000 vaccinations. This cohort study has not been included in the analysis as the vaccine studied is no longer in use.

Serious adverse events ‐ Guillain‐Barré syndrome ‐ case‐control studies (Analysis 07)

In an analysis performed using the mean of unadjusted data relative to six data sets, exposure to monovalent H1N1 pandemic inactivated vaccine resulted in an apparent statistically significant association with Guillain‐Barré syndrome onset when administration took place within six weeks before symptoms occurred (odds ratio (OR) 2.22, 95% CI 1.14 to 4.31, Analysis 7.1.1). Thus, it should be taken into account that only one out of the six data sets showed a statistically significant association between vaccine exposure and Guillain‐Barré syndrome onset (bb Dieleman 2011e). When we performed a sensitivity analysis excluding this data set from the pooled estimate, the result was no longer significant. When the analysis was performed considering that vaccine exposure occurred at any time before disease onset, there was no significant association (Analysis 7.1.2).

The analyses performed by pooling authors' estimates adjusted for several confounders (i.e. receipt of other vaccines, family history of autoimmune diseases, physician consultation during the previous year and use of antibiotic, antiviral or antipyretic agents) do not show a statistical association for exposure within six weeks (Analysis 7.2.1) before disease onset or for exposure at any time (Analysis 7.2.2).

Data from one other case‐control study confirm that immunisation with seasonal inactivated vaccine is not significantly associated with the onset of Guillain‐Barré syndrome within six weeks after inoculation (bb Galeotti 2013) (Analysis 7.3).

Serious adverse events ‐ demyelinating diseases ‐ cohort studies (Analysis 08)

In one cohort study the authors tried to assess whether there is an association between exposure to inactivated trivalent seasonal influenza vaccine during pregnancy and several pathologies (e.g. Guillain‐Barré syndrome, demyelinating diseases, immune thrombocytopaenic purpura) within six weeks after immunisation. Unadjusted estimates were calculated for an association with demyelinating diseases by using the number of cases observed among exposed and unexposed hemi‐cohorts and indicate that there is no association (Analysis 8.1.2).

One cohort study assessed the safety of the H1N1 vaccine. No statistical association was found between vaccination with H1N1 monovalent pandemic vaccine and demyelinating diseases.

Serious adverse events ‐ demyelinating diseases ‐ case‐control studies (Analysis 09)

An association between exposure to seasonal inactivated vaccine and demyelinating diseases (including both multiple sclerosis and optic neuritis case definitions) in a healthy adult population was not statistically significant when we pooled unadjusted data from four case‐control studies (OR 0.96, 95% CI 0.79 to 1.17) (Analysis 9.1). Also, when we analysed adjusted data for each of the case definitions separately, the estimates remained non‐statistically significant for multiple sclerosis (Analysis 9.2) and for optic neuritis (Analysis 9.3).

Serious adverse events ‐ immune thrombocytopenic purpura ‐ cohort studies (Analysis 10)

One cohort study aimed to assess whether there is an association between exposure to inactivated trivalent seasonal influenza vaccine during pregnancy and several pathologies (e.g. Guillain‐Barré syndrome, demyelinating diseases, immune thrombocytopaenic purpura) within six weeks after immunisation. Neither the unadjusted (Analysis 10.2.2) nor adjusted estimates (Analysis 10.1.2) for an association with immune thrombocytopenic purpura were statistically significant.

Serious adverse events ‐ immune thrombocytopenic purpura ‐ case‐control studies (Analysis 11)

Data analysis of two case‐control studies (bb Garbe 2012; bb Grimaldi‐Bensouda 2012) did not show a statistically significant association between immune thrombocytopaenic purpura and seasonal influenza vaccine in any of the time frames considered (i.e. less than two months, six or 12 months between immunisation and disease onset), or when the data were pooled together (Analysis 11.2). The same conclusions could be drawn when analysis was performed by using estimates adjusted for confounders (Analysis 11.1) and are further confirmed by the fact that a sensitivity analysis carried out by using either a random‐effects or fixed‐effect model did not change them in any way. It should be observed that no data sets included in this comparison, with the exception of bb Garbe 2012, showed a statistical association between disease and influenza vaccination. It is possible that the ages of the participants (cases and controls) were different in these two studies and that some elderly participants could have been included. Unlike bb Grimaldi‐Bensouda 2012, the case‐control study (bb Garbe 2012) considered as exposed those cases that were immunised up until 28 days before immune thrombocytopaenic purpura onset.

Vaccines for the 1968 to 1969 (H3N2) influenza pandemic (Analyses 12 to 16)

Five studies yielded 12 data sets (aa Eddy 1970; aa Mogabgab 1970a; aa Mogabgab 1970b; aa Sumarokow 1971; aa Waldman 1969a; aa Waldman 1969b; aa Waldman 1969c; aa Waldman 1969d; aa Waldman 1972a; aa Waldman 1972b; aa Waldman 1972c; aa Waldman 1972d). As one would expect, vaccine performance was poor when the content did not match the pandemic strain (Analysis 12.1; Analysis 12.2). However, one‐dose or two‐dose monovalent whole virion (i.e. containing dead complete viruses) vaccines achieved a VE of 65% (95% CI 52% to 75%) protection against ILI (NNV 16, 95% CI 14 to 20), a VE of 93% (95% CI 69% to 98%; NNV 35, 95% CI 33 to 47) protection against influenza and a VE of 65% (95% CI 6% to 87%) with NNV 94 (95% CI 70 to 1022) against hospitalisation (Analysis 13.1; Analysis 13.2; Analysis 13.3).

Approximately half a working day and half a day of illness (Analysis 13.5; Analysis 13.6) were saved but no effect was observed on pneumonia (Analysis 13.4). All comparisons except for ILI are based on a single study (Analysis 13.4). The large effect on ILI is coherent with the high proportion of these illnesses caused by influenza viruses in a pandemic (i.e. the gap between the efficacy and effectiveness of the vaccines is narrow). Aerosol polyvalent or monovalent vaccines had a modest effect.