ATTRIBUTABLE RISK ESTIMATION: AN OPEN ISSUE

To the Editor:

We read with interest the article "Risk factors and outcome of nosocomial infections: Results of a matched case-control study of ICU patients" by Girou and colleagues (1) reporting a matched case-control study on risk factors and outcome of nosocomial infection in the intensive care unit (ICU). They reported a 44% estimated mortality attributable to nosocomial infection. This is quite an important clinical research question with wide public health consequences. However, many methodological points of this paper are worth noting that should broadly limit its conclusions.

As expressed in the title, the study had two main objectives: estimating the risk factors for, and the outcomes of, nosocomial infection in ICU. However, the second part of the title is misleading. Indeed, although the study of risk factors for nosocomial infection used a case-control design, the assessment of attributable risk of death in ICU patients with nosocomial infection actually used a matched exposed-unexposed cohort.

The attributable (ICU or hospital?) mortality was defined as the risk difference between infected and noninfected patients; it was estimated by subtracting the observed crude mortality rate of the unexposed patients from that of the exposed ones. Besides the inconsistent usage of this definition (which actually refers to the risk difference in exposed patients) it has been widely shown that the unadjusted estimation of attributable risk is biased, so that adjusted estimation on confounding factors and secondary exposure is mandatory. This study showed perfectly that severity of illness is a major confounding factor, as severity of illness is a risk factor for nosocomial infection and is evidently related to patient death. However, no adjustment was performed in the analysis. The lack of such adjustment could have been compensated for by relying on the study design, i.e., infected and noninfected patients were matched on three possible confounders, age (± 5), APACHE II (± 5) on admission, and length of stay before infection, that had to be at least equal to the interval, in cases, from admission to first infection. Nevertheless, as we discussed above, secondary exposures and confounders, such as the reported differences within the three first days of ICU admission between matched exposed and unexposed patientsreflecting the fact that patients with increasing severity are at high risk of developing nosocomial infectionwere not taken into account. Perhaps it is likely that such an adjustment would have erased the observed difference in risk and allowed a better explanation of the "interrelationships between underlying disease, severity of illness, therapeutic activity and nosocomial infection."

Finally, analyses of other outcome measures appear questionable. For instance, given the previous reports of time distribution of nosocomial infections in ICU, it is likely that many episodes would have occurred within the first week of admission, thus Figure 1 in their article, which displays the (mean?) severity score values over time, may be difficult to interpret.

SYLVIE CHEVRET

JEAN-FRANÇOIS TIMSIT

Departement de Biostatistique et

  Informatique Medicale

University of Paris

Paris, France

	References

1. Girou, E.. 1998. Risk factors and outcome of nosocomial infections: results of a matched case-control study of ICU patients. Am. J. Respir. Crit. Care Med. 157: 1151-1158 [Abstract/Free Full Text].

From the Authors:

Drs. Chevret and Timsit point out that our study did not totally exclude confounders and secondary exposures as factors influencing the prognosis of ICU patients with or without nosocomial infections. If we correctly interpret their theoretical and methodological concerns, the only ways to avoid such confounding variables are to pair cases and controls at the time of death or to experimentally investigate ventilated animals to ensure the absence of underlying disease(s). Our clinical investigation aimed to analyze the ICU course of severely ill patients from admission to discharge or death, taking into account modifications of clinical status over time, and major confounders well known to influence survival: age, severity of illness scores, and duration of ICU stay (i.e., duration of exposure to risk). Thus, with the data presented in our study, readers were able to interpret interrelationships between underlying disease, severity of illness, and level of therapeutic activity and their variations during ICU stay, and nosocomial infections (including type of infection, number of infections, and time of infections).

The 44% difference between mortality rates of cases (i.e., infected or exposed patients) and controls (i.e., noninfected or unexposed patients) can be attributed to the type of underlying disease and the level of physiological reserves, severity of illness, efficacy of therapeutics, accuracy of diagnostic strategies, as well as the onset of various unexpected events complicating the course of patients admitted to ICU, including nosocomial infections. By definition, similar admission APACHE II scores, confirmed by similar SAPS, Glasgow coma scores, McCabe scores, TISS scores, and ODIN scores reflect similar expected death rates for cases and controls. Nevertheless, we observed a fourfold higher actual mortality rate in patients who developed nosocomial infections as compared to control noninfected patients. In our opinion, the probability appears to be very low that the variable tested in our studynosocomial infection, by design differently distributed between cases and controlswas the only variable not to play a significant prognostic role when variables equally distributed with the pairing process had major contributions to the outcomes of these patients.

In Figure 1 in our article, NI is defined as the day of the first nosocomial infection regardless of the time between Day 3 and discharge at which it occurred. The negative and positive days surrounding NI are determined as a function of NI itself (i.e., before and after the event). D1, D2 and D3 are also fixed dates, as nosocomial infection is defined as occurring after at least 48 h in the ICU. However, we now see that we did not explain n, which is the number of patients present in the ICU on that day. Thus, NI developed after Day 7 in only 18 patients.

JEAN-YVES FAGON

EMMANUELLE GIROU

Service de Réanimation Médicale

Hôpital Broussais

Paris, France