Factors affecting the development of ventilator-associated pneumonia (VAP) [110] have been assessed with several methods, including multivariate analysis. Most studies used logistic regression [27], but several have used failure-time analysis to compare the interval from the start of mechanical ventilatory support to the onset of VAP [2, 810]. Because the duration of ventilatory support affects the VAP rate [11], this ventilation period should be examined when the VAP rate is analyzed. Logistic regression can be used to examine only those variables that are independent of the effect. However, the total ventilation period is always affected by the presence of VAP, although VAP itself is also affected by the ventilation period (before the onset of VAP). Therefore, the “clinically required ventilation period,” excluding the effect of VAP, should be used instead of the total ventilation period. This clinically required ventilation period is the imaginary ventilation period during which a patient would have required ventilatory support if they had not had VAP; this period might be longer than the ventilation period before the onset of VAP. However, this clinically required ventilation period has not been predictable, and therefore studies using logistic regression have been biased.

Comparison of the time before the onset of VAP, called failure-time analysis in the present study, is another approach for assessing factors that affect VAP. Although this method solves the problem of the clinically required ventilation period, it is rarely used [2, 810], possibly because of difficulties in assessment, especially in multivariate analysis. Many multivariate methods can use only parametric models, and even the Cox proportional hazard model is an approximate model for nonparametric data. In addition, right-censored (suspended) data complicate both assessment and calculation.

We attempted to resolve these problems by using rate analysis and failure-time analysis and comparing the results.

Patients and methods

Study location and patients

The data were retrospectively collected from April 2009 through May 2010 in intensive care units (ICUs) of university hospitals in Tokyo, Japan. The criterion for patient inclusion was having received mechanical ventilatory support in an ICU. Exclusion criteria were absence of tracheal intubation, tracheal intubation after onset of pneumonia, and patients belonging to departments in which no VAP occurred. Because VAP is usually defined as pneumonia developing more than 48 h after the start of mechanical ventilation [12, 13], pneumonia developing 48 h or less after the start of mechanical ventilation was distinguished from VAP (with no episode).

Data collection

Chest X-ray films were routinely obtained for each patient receiving mechanical ventilatory support and were examined by two board-certified members of the Japanese Respiratory Society who made diagnoses after consulting one another. All enrolled patients were examined as to whether they belonged to the categories defined later. Data were collected retrospectively before and after discharge from the ICU. The latest clinical examination findings and latest laboratory data, obtained from routine blood analysis, were collected before intubation. Because intubation was usually (94.9%) necessitated by surgical procedures, most data were collected before surgery.


The criteria for diagnosing VAP were standardized [12, 13] and included (1) chest X-ray film indicative of pneumonia; (2) body temperature greater than 38°C, leukocyte count greater than 10,000/mm3 or less than 4,000/mm3, or an increase in the C-reactive protein level; (3) low oxyhemoglobin saturation or an increase in oxygen need; and (4) increased production of sputum containing possibly pathogenic microorganisms. Therefore, the rates of ventilator-associated tracheobronchitis [14] with congestive heart failure, acute respiratory distress syndrome, and interstitial pneumonia might have been overestimated. Microbiological specimens were always obtained before the administration of new antibiotics, as soon as possible after clinical or radiologic abnormalities were found. Only the first episode of VAP was analyzed.

Statistical analysis

The software program JMP 8.0.2 (SAS Institute, Cary, NC, USA) was used for analyses. In both univariate and multivariate analyses, logistic regression was used to estimate the VAP rate (rate analysis), and the Cox proportional hazard model was used to estimate ventilation time before VAP (failure-time analysis). Logistic regression of single discrete variables was equivalent to the likelihood ratio test of the chi-square test. Because JMP 8.0.2 could not be used to analyze time-dependent covariables [15], a time-independent hazard model was used. Odds ratios from logistic regression and hazard ratios from the Cox proportional hazard model were compared and checked for differing tendencies. Because we failed to determine the clinically required ventilation period, for multivariate rate analysis we instead used the total ventilation period and the ventilation period before VAP. We compared these two period models and assumed that coincident factors in them might be the same as the true factors that would be calculated using the clinically required ventilation period if that period were detectable.

For failure-time analysis, a parametric model was also used. After comparing Frechet, Weibull, log-normal, log-logistic, exponential, and other distributions, we performed parametric multivariate analysis using the most suitable Frechet distribution. Because of the small sample size, we failed to assess interactions of all factors, but we did assess interactions after significant factors remained. Interaction effects were represented as nested effects. For assessing differences between the total ventilation period and the ventilation period before VAP, the one-tailed Wilcoxon signed-rank test was performed.


Data were collected for all 325 patients who had received ventilatory support. Eleven patients who had pneumonia or interstitial pneumonia were excluded, and 39 patients who had been admitted by ten departments that had no episodes of VAP were excluded. Consequently 275 patients, including 12 patients with VAP, were included in this study. Fourteen patients who did not undergo surgery were also included. No patients had pneumonia within 48 h of the start of ventilatory support.

Characteristics of patients and results of univariate analyses are summarized in Table 1. Operation times are not given for the 14 patients who did not undergo surgery. In patients with VAP, the total ventilation period was significantly longer than the ventilation period before VAP (P = 0.0002 by one-tailed Wilcoxon signed-rank test). On univariate rate analysis, statistically significant values were seen for the Glasgow Coma Scale, the APACHE II score, operation time, total ventilation time, and nonfeeding days. However, univariate failure-time analyses revealed no statistically significant values. Most odds ratios and hazard ratios were similar, but some were opposite. For example, in the emergency category the odds ratio for emergency admissions (1.435) indicated an increased rate of VAP. In contrast, in the emergency category the hazard ratio for emergency admissions (0.3298) indicated the development of VAP was delayed.

Table 1 Examined factors and univariate analysis for ventilator-associated pneumonia (VAP) rate and VAP failure time

Results of multivariate rate analyses are summarized in Tables 2 and 3. With both methods, operation time and body mass index (BMI) had statistically significant effects, indicating that these factors might be significant in a model using the “clinically required ventilation period.” The total ventilation (period) in Table 3 did include the VAP effect; therefore, this factor should not be interpreted in this model. In addition, the ventilation period before VAP was not significant in Table 2, indicating that we cannot conclude the clinically required ventilation period had an effect in this model.

Table 2 Multivariate rate analysis for before VAP
Table 3 Multivariate rate analysis for before and after VAP

Results of multivariate failure-time analysis are summarized in Tables 4 and 5. With both the Frechet distribution model and the Cox proportional hazard model, sex and female BMI had statistically significant effects. However, emergency was only significant in Table 5, indicating that the Frechet distribution model might be more suitable than the Cox proportional hazard model.

Table 4 Multivariate failure-time analysis from the Cox proportional hazard model
Table 5 Multivariate failure-time analysis from the parametric model of the Frechet distribution


We have used two-way analysis—rate analysis and failure-time analysis—for predicting VAP. To our knowledge only one similar method has been reported previously [2], although the characteristics of the method, interpretation of the method, or how to handle the clinically required ventilation period were not described Therefore, we discuss these problems next.

In rate analysis, we missed the interaction of BMI and sex, because it was not statistically significant. However, because failure-time analysis showed these factors to be significant, we reexamined this interaction and found the nested effect had the same tendency. Therefore, this two-way method might reduce this type of error.

The discrepancy in the “emergency” effect on univariate analysis was caused by a time-dependent bias. Emergency admission to the ICU prolonged the time before VAP developed but also prolonged the ventilation time. Multivariate rate analysis showed that patients emergently admitted to the ICU tended to have a lower VAP rate when the ventilation duration was adjusted for, but the difference was not statistically significant and was, therefore, ignored. However, multivariate failure-time analysis revealed that the “emergency” effect significantly prolonged the time before VAP. In contrast, multivariate failure-time analysis did not identify the effect of “operation time,” although multivariate rate analysis showed it to be statistically significant. These differing results had the same tendency but might agree if there were a large amount of data. The two-way analyses described in this study might fail less frequently to identify these phenomena.

These two methods are not rigorous. Rate analysis failed to identify the clinically required ventilation period as a significant factor affecting the rate of VAP. Failure-time analysis is not rigorous because it is only an approximate method [16]. Therefore, these methods are not rigorous themselves, but the use of both methods together might improve reliability. In addition, by comparing these methods, we can choose the more suitable one, if only a single method is to be used.

Discrepant factors in this two-way method suggest the existence of hidden factors that could not be identified. For example, emergently admitted patients had a longer ventilation period (median, 6 days; interquartile range, 3–15 days) than did other patients (median, 2 days; interquartile range, 1–3 days). Because of the longer ventilation period, emergently admitted patients tended to have a higher rate of VAP. To avoid VAP in emergently admitted patients, the ICU staff would exercise greater care, which might prolong the time before VAP. Even if the effect of greater care did not exist, a hidden factor would be causing this unexplained prolongation of time before VAP.

Both the ventilation period and the operation time not only mean the time itself but also represent the severity of the disease, for the following reason. If a patient has more severe disease, operation time [5, 17] or ventilation period [11, 18, 19] would tend to be longer. Perhaps the failure to estimate the “clinically required ventilation period” helped us to find the operation time effect with the severity effect. Operation time can be determined much more easily than can clinically required ventilation period or severity. Therefore, accurate observation of operation time would be more useful than uncertain estimates of clinically required ventilation period or severity. The two-way method might allow us to find a more suitable factor.

Some previous studies excluded episodes of VAP developing after less than 48 h of ventilatory support [3, 4, 9] because the definition of VAP also excludes such episodes. However, the focus of the present study was the method of analysis rather than factors leading to VAP. Therefore, we gave priority to including many subjects instead of excluding subjects with biased information. If we had an adequate sample size, we could exclude episodes occurring with a ventilation period of less than 48 h. Unfortunately, we could not examine every factor used in this study without episodes of VAP developing before a ventilation period of 48 h. Adding analyses excluding these early episodes revealed the same tendency as in multivariate analysis; therefore, we also interpreted the results of analyses with early episodes.

Another problem was encountered before interpretation. We had examined levels of brain natriuretic peptide (BNP) before admission as a comorbidity factor representing cardiac failure. However, BNP is rarely examined, except by departments of cardiovascular surgery. In addition, in patients with high BNP levels, cardiac failure is usually treated with surgery. Unfortunately, we examined BNP and other factors reflecting cardiac function before surgery rather than after surgery. Although BNP measured before admission had a bias that appeared significant in the analyses models, we excluded 172 sampled, biased, and clinically unimportant measurements of BNP before admission.

An earlier study [6] had found that higher BMI increased the rate of VAP, but our data show that higher BMI decreases the rate of VAP. These results appear contradictory but may be explained by a difference in distribution. BMI usually shows a J-shaped correlation with many factors, such as mortality [20] and infection rate [21]. The earlier study found that 37.3% of patients had a BMI <25 kg/m2, but the present study found that 82.5% of female subjects had a BMI <25 kg/m2. Therefore, a J-shaped correlation of BMI with VAP will explain this different tendency.

Few earlier studies have considered the correlation of operation time with VAP [5]. Our data suggest that operation time is associated with the rate of VAP, but our failure-time analysis failed to find this association. More data are needed to examine this effect. When a patient who has more severe disease undergoes surgery, operation time tends to increase. Therefore, operation time would correlate with disease severity. In addition, operation time is easily determined. Therefore, further investigation of the effect of operation time on VAP is warranted.

We have performed the same type of two-way analysis for postchemotherapeutic febrile neutropenia in patients with hematological malignancies, using previously reported data [22]. Because the problem of clinically required ventilation period was not involved, only rate analysis was performed [2325]. We found that patients with acute myelogenous leukemia had a higher fever rate but their fever occurred later. The discrepancy in univariate analysis might be explained by time-dependent bias, and the peculiar prolongation of fever in male patients on multivariate analysis suggests that staff efforts, such as preparation of sterile food, reverse isolation in private rooms, and antibiotic prophylaxis, prolonged the fever. We speculate that such artificial efforts might cause this type of discrepancy.

We analyzed the accident rate and the time before accidents, which, in this case, was the development of VAP. These two variables are similar but differ in some ways, such as emergency admission. Unknown factors might be obscured by such differences. However, distinguishing a true difference from a technical difference can be difficult and requires keen clinical or situational insight. We believe that if an accident occurs frequently in a group, efforts would be made to avoid the accident. Such efforts would prolong the time before an accident occurs and reduce the accident rate. However, sometimes the accident rate in a given group, although lower, remains higher than that in another group, whereas the time before an accident becomes longer than in the other group. This situation may create a time-dependent bias and peculiar prolongation of the time before an accident. Therefore, an artificial effect of this type would, in theory, be detected with this two-way method. This two-way method has rarely been used but could be used for various types of study. This method is somewhat complicated and can easily be misinterpreted, but we hope that many researchers will use and verify this two-way method.