We restricted our analyses to patients with meningitis, including those with stroke caused by presumed meningitis. Cases of parameningeal (e.g., epidural abscess) or peripheral joint infection were not included in this analysis. To evaluate the effect of public health actions on the size of this outbreak, we sought to determine numbers of 1) MPA injections averted, 2) cases of meningitis or stroke averted, and 3) deaths caused by meningitis or stroke averted. Information about volume of MPA recalled was unavailable, so we estimated the recalled volume from data in NECC MPA shipping records. To estimate cases averted, we calculated attack rates in 2 steps. First, we modeled attack rates for patients with single injections. For patients with multiple injections, we conducted a Monte Carlo simulation that assigned a single responsible injection for each patient in each simulation. This simulation also provided estimates of uncertainty. To estimate deaths averted, we applied the case-fatality rate (CFR) for case-patients before most public health actions had occurred for the case-population at risk (i.e., number of case-patients diagnosed after October 4 and still alive 60 days later and estimated number of case-patients predicted to have occurred after October 4 if no recall had occurred).

Case definitions developed during the outbreak response were used in this analysis (9). A case-patient was defined as a person who, after May 21, 2012, received an epidural or paraspinal MPA injection from an implicated lot and subsequently had either meningitis of unknown etiology or posterior circulation stroke without cardioembolic source and without a documented normal cerebrospinal fluid (CSF) profile. Meningitis was defined as >5 leukocytes/mm³ in CSF from a person with compatible symptoms (e.g., headache, stiff neck). Date of diagnosis was defined as the date when an initial lumbar puncture yielded CSF that met the meningitis case definition or the reported date of stroke diagnosis for stroke cases. Cases were defined as laboratory confirmed if histopathologic, microscopic, culture, or molecular evidence of a fungal pathogen was present and associated with the clinical syndrome. Inclusion in this analysis was limited to cases reported on or before the final case count published on October 23, 2013.

NECC shipping records provided information on date of shipment, volume, and lot number of MPA shipped to each of the 75 injection facilities that received contaminated lots of MPA. To estimate volume of steroid recalled, we modeled the MPA use rate at each injection facility. For the 52 facilities that received >1 shipment of implicated MPA, we estimated a per-facility MPA use rate on the basis of reorder frequency and shipment volumes. This approach could not be used for 23 facilities that had a single shipment (i.e., they placed no reorders). For those facilities, we developed a simple linear regression model to estimate use rate with average volume of MPA as the predictor (Technical Appendix). We assumed that use of NECC MPA at each injection facility began on the first business day after a shipment date, that use occurred at the facility only on business days, and that use ceased completely on September 27, 2012. These methods enabled us to estimate weekly volume of NECC MPA used across all 75 clinics.

To estimate the volume recalled, we subtracted the total volume of estimated use before September 27 from the total volume shipped. To estimate the number of MPA exposures potentially averted, we divided the total volume of MPA recalled by 1 mL/injection volume (i.e., amount of MPA used per injection for 80% of patients).

Previously published data suggested that attack rates varied by lot number and age of the vial, with age being measured as length of time from vial production to injection date (1). Consequently, we anticipated that we would need to estimate separate attack rates for each of the 3 implicated MPA lots to account for vial age. To account for changes in attack rates by vial age, we calculated the attack rate for each calendar week during which injections took place and produced separate estimates for each lot of MPA. For case-patients with 1 injection, weekly lot-specific attack rates were calculated as follows: number of case-patients receiving an injection of lot X during a given week divided by estimated volume of lot X used that week (denominator was obtained from the MPA use rate analysis). A Poisson distribution was then fitted to the observed data to estimate the attack rate for each week during which injections were given (Technical Appendix). The model’s covariates were week of injection, lot number, and interaction between week and lot variables, selected because these variables showed evidence of modifying results.

Next, for case-patients with >1 injection, we developed a probabilistic model that used a Monte Carlo simulation to randomly assign a single injection as responsible for each case-patient’s infection (Technical Appendix). A total of 100 simulations were performed; for each simulation, the probability of injection assignment was derived from the predicted weekly lot-specific attack rates among case-patients with only 1 injection. A Poisson regression model, using the same covariates as those used in the model for case-patients with 1 injection, was then fit to the 100 simulated datasets to produce a distribution of predicted number of cases (Technical Appendix). To capture additional uncertainty, 100 additional simulations of the predicted number of cases were generated so that 10,000 estimates of the predicted number of cases were generated for each week and lot combination.

From this distribution, we then summed the median number of cases predicted for each week and lot combination after the recall date to obtain an estimated total number of cases averted. Finally, 95% CIs were calculated to capture uncertainty in our estimates (Technical Appendix). Extrapolation of weekly attack rates for lots 05 and 06 were continued until the model predicted that the volume of MPA in those lots had been completely used (Figure 1). However, lot 08 was in use for only 5 weeks before the recall on September 26, 2012. Thus, for this lot, we limited the extrapolation of attack rates and cases averted to 3 weeks after the recall (until October 14, 2012).

To estimate the number of deaths averted, we first calculated the 60-day CFR for patients given a diagnosis early in the outbreak (on or before October 4), before widespread publicity and posting of clinical treatment guidelines. This 60-day CFR was then applied to 2 groups of case-patients to estimate deaths averted: 1) patients given a diagnosis after October 4 who were still alive 60 days after diagnosis and 2) averted cases predicted to have occurred after October 4 without a recall.

The 60-day CFR was plotted by week of diagnosis to examine mortality trends during the course of the outbreak. To understand whether these mortality trends were associated with changes in case-patient clinical characteristics, we compared measures of disease severity and treatment rapidity for case-patients given a diagnosis on or before October 4, 2012 to those given a diagnosis after that date. A χ² test (dichotomous variables) and the Wilcoxon rank-sum (continuous variables) test were used. We estimated the survival function of time from diagnosis to death by using a Kaplan-Meier estimator. The log-rank test was used to compare survival of the 2 groups.

Data on case-patient clinical characteristics were collected from a standardized case report form created for this outbreak response. States reported case-patient deaths and dates of death routinely to CDC; thus, deaths occurring after completion of a case report form were captured systematically. The collection of patient-level data was deemed a public health emergency and was not subject to ethics review or informed consent procedures.

NECC records indicated that 17,675 vials comprising 29,641 mL of the implicated MPA lots were distributed to 75 facilities in 23 states. Of the 11,773 mL of lot 05 shipped, we estimated that none was recalled; of the 10,847 mL of lot 06 shipped, we estimated that 132 mL (1%) was recalled; and of the 7,021 mL of lot 08 shipped, we estimated that 3,018 mL (43%) was recalled (Table 1) (Technical Appendix). A total estimated 3,150 mL of MPA was shipped but not administered because of the recall. Using an estimated per-injection volume of 1 mL, an additional 3,150 injections with the implicated lots of MPA could have occurred if all 3 lots of contaminated MPA had not been promptly recalled (Table 1).

Of the 391 case-patients, 370 (95%) had injection data available for calculating attack rates. Of the 370 patients, 221 (60%) had 1 injection, 116 (31%) had 2 injections, 32 (9%) had 3 injections, and 1 (<1%) had 4 injections. Figure 1 shows observed and estimated weekly lot-specific attack rates in patients for each of the 3 contaminated lots of MPA.

For lot 05, predicted weekly attack rates increased from 1.9 (95% CI 0.8–4.7) cases/1,000 mL MPA administered during the first week of injections to 25.8 (95% CI 7.7–81.3) cases/1,000 mL MPA administered during the last week of injections; for these predictions, we assumed that injections continued until all vials were used. Lot 06 had predicted weekly attack rates that increased from 1.6 (95% CI 0.8–3.1) cases/1,000 mL MPA administered during the first week to 418.6 (95% CI 226.8–768.1) cases/1,000 mL MPA administered during the final week of injections (with the assumption that injections continued until all the vials were used). Lot 08 had predicted weekly attack rates that increased from 5.5 (95% CI 1.2–23.0) cases/1,000 mL MPA administered during the first week to 57.3 (95% CI 10.7–302.8) cases/1,000 mL MPA administered during the third week (October 14, 2012) after the recall notice. For all 3 lots, predicted attack rates rose as time from the medication production date increased (Figure 1). We estimate that without the recall of the 3 lots of MPA, 153 (95% CI 61–467) additional cases of meningitis would likely have occurred in persons exposed between September 26, 2012, and October 14, 2012 (Table 2).

Of the 389 patients with available detailed data, 40 died <60 days of diagnosis (60-day CFR 10%). Of the 82 patients given a diagnosis on or before October, 23 (28%) died <60 days of diagnosis compared with 17 (6%) of the 307 patients given a diagnosis after October 4 (p<0.0001), an absolute risk reduction of 22% (Table 3). Of the 110 patients with laboratory-confirmed infection, 19 died <60 days of diagnosis (60-day CFR 17%). Of the 33 patients given a diagnosis on or before October 4, 13 (39%) died <60 days of diagnosis, compared with 6 (8%) of 77 who had a diagnosis after October 4 (p<0.0001) (Table 3).

Kaplan-Meier analysis and the log-rank tests showed that patients given a diagnosis after October 4 had improved overall survival compared with those given a diagnosis on or before October 4 (p<0.0001) (Figure 2). Patients given a diagnosis after October 4 also had lower CSF median leukocyte count (29 cells vs. 1,064 cells; p<0.0001); lower median CSF protein (71 g/dL vs. 117 g/dL; p<0.0001); higher median CSF glucose (55 mg/dL vs. 38 mg/dL; p<0.0001); and fewer symptoms when care was sought (median 4 vs. 5; p<0.0001) (Table 3). For the 240 patients with documented receipt of antifungal treatment, patients given a diagnosis after October 4 were more likely to receive antifungal drugs <48 hours of diagnosis (84% vs. 59%; p = 0.0006) (Table 3). The 60-day CFR by week of diagnosis was 50%–100% from August 27 through September 30; the CFR dropped substantially during the first week of October, to 3%–7% (Figure 3).

We applied the 60-day CFR (28%) of the 82 patients given a diagnosis of meningitis on or before October 4 to that of the 290 patients whose meningitis was diagnosed after October 4 and who were alive at 60 days (Table 3); we also applied that CFR to the 153 (95% CI 61–467) cases estimated to have been averted. With these calculations, we estimated that, without public health actions and early diagnosis and intervention by clinicians, 124 (range 98–211) additional deaths from meningitis or stroke may have occurred through October 14, 2012 (Table 2).

Extrapolation of the amount used from lot 08 suggests that, in the absence of a recall, its use would have continued until the week of December 2, 2012 (Figure 1). Estimates of attack rates and cases averted beyond October 14, 2012, are unreliable because of the short amount of time that lot 08 was in use before the recall. However, if lot 08 injections had continued through December 2 and if we assume an exponential rate of increase in attack rates (as was applied to lot 06), during the final week of injections, the attack rate for lot 08 might have been as high as 768.8 cases/1,000 persons exposed. Under these same assumptions, an additional 169 cases (total estimated cases = 322) and an additional 47 deaths (total estimated deaths = 171) would have occurred if lot 08 had been used through December 2.