Antipsychotic-Induced Weight Gain: A Comprehensive Research Synthesis

Studies were coded by one investigator (J.L.M.) and spot-checked by one of two other investigators (M.H. or D.B.A). When a discrepancy was found (a fairly rare event), the coders met to discuss and resolve the discrepancy.

The mean and standard deviation of weight change and the size of each group were the three essential pieces of information needed from the studies. In many cases, these data were reported directly in the article and simply recorded. However, in other cases, they were not. In this latter situation, one of several approaches was taken in the following order of preference.

1. Missing means, standard deviations, or sample sizes were directly calculated by using other information available in the article (for example, t, F, or p values) and standard statistical formulas (20).

2. If the article was published in 1990 or later, we attempted to contact the authors for more information.

3. Two other procedures were used to estimate (rather than directly calculate) the necessary statistics. One method was used when data were presented in “binned” categories (e.g., “Ten percent of the patients gained no weight, 30% gained 0–5 pounds, 40% gained 5–15 pounds, and 20% gained more than 15 pounds”). In these situations, by using the categories and the proportions of subjects in each category, the missing mean and/or standard deviation was estimated by maximum likelihood methods; that is, we simply found the estimates of the means and the standard deviations that maximized the likelihood of the observed data by using the normal distribution likelihood function (21). The second method was used when the standard deviation was not reported but the range was (e.g., “Weight change ranged from –4 kg to +15 kg”). In this case we adapted the approach of Tippett (22), who published tables that, given the sample size, provide the expected ratio between the sample range and the standard deviation. Using Tippett’s method, we estimated the standard deviation.

4. If only the standard deviation was missing, it was estimated as the square root of the weighted average variance across all other studies where the weights used were the sample sizes in each study. It was necessary for a standard deviation to be available in order to estimate the variance of the mean for each study, so that the inverse of this variance could be used as a weighting factor in subsequent analyses.

Finally, if none of these methods could be used to estimate the mean, standard deviation, and size of a study sample or a subgroup within a study, that study or subgroup was excluded from further consideration in the formal statistical meta-analysis. The total number of studies yielding usable data was 81. These studies yielded a total of 418 estimates of weight change in some antipsychotic drug condition or nondrug control condition. Of these 418 data points, 96.7% of the means, 69.6% of the standard deviations, and 100% of the numbers of study subjects were obtained by transcription or calculation, and the remainder by some form of estimation or imputation. Table 2 shows the mean and range of time on medication (in weeks) for the observed data points on each drug.

Before the statistical meta-analysis was conducted, a verbal overview was done, because several articles provided descriptive data on weight change that could not be included in the quantitative analysis but nevertheless offered some information. Key quotations that characterized the effect of the drugs in question were extracted from such articles.

Statistical analyses were conducted with SPSS, version 7.5 (23). The effects of antipsychotic drugs were analyzed separately for each drug, since preliminary analyses indicated marked differences among the specific compounds in terms of their effects. Because most studies did not include a placebo comparison group, the effect size we used was the raw weight change from baseline to posttreatment. Only 18 studies included placebo comparisons. By using the pretreatment-to-posttreatment weight change in all studies, we were able to make full use of all of the available data.

Since there were 19 different drugs/conditions (including placebo; nonpharmacologic, nonplacebo control; and polypharmacy), 19 separate analyses were conducted (one for each condition). For each condition we attempted to calculate the weighted mean weight change and standard error based on both a fixed effects model (24) and a random effects model (24). Although both the fixed and random effects estimates are presented in the tables, only the random effects estimates are discussed in the text, given the significant heterogeneity present for most compounds (see the Results section).

For each drug, when sufficient data (i.e., six or more data points) were available, we regressed mean weight change on standardized drug dosage and length of treatment. One older, poorly controlled study (25) was eliminated because it was an outlier, and its exceptionally long follow-up of 11 years caused it to act as a leverage point (26); all of the other follow-ups were less than 200 weeks long. These regressions were conducted as weighted least squares multiple regressions, where the weights were equal to the inverse of the variances of the dependent observations. To more reasonably compare drugs by controlling for different dosage levels, we calculated standardized doses by dividing the actual doses used in the studies by the midpoint of the recommended dose range and taking the natural log of the resulting ratio. (Although we adhered to this procedure for all drugs in the interest of consistency, it is possible that in some cases, the midpoint of the recommended dose range may not have been the best ­estimate of the standard dose. Therefore, for the atypical antipsy­chotics, haloperidol, and thioridazine [the most commonly used drugs], we conducted a sensitivity analysis by recomputing the results. We replaced the standardized dose first with the typical dose in chlorpromazine equivalents according to APA’s Practice Guideline for the Treatment of Patients With Schizophrenia [27] and second with the average dose used in clinical settings as reported in the peer-reviewed literature.) Recommended dose ranges were obtained from the appendix of a consensus report (28), the Physicians’ Desk Reference, or the drug manufacturer. The regression equation we used was Δ_kg=β₀ + β₁(weeks – 10) + β₂(weeks – 10)² + β₃(D) + β₄(D)² + e, where Δ_kg is weight change in kilograms, the βs are parameters to be estimated, weeks is number of weeks of treatment, D is the standardized dose calculated as described above, and e is an error term. In this equation, β₀ is a direct estimator of weight change at 10 weeks at the standard dose. For placebo, nonpharmacologic control, and poly­pharmacy, dosage information was not included in the regression.

Using the aforementioned equation, we estimated the weight-promoting effects of each drug at the midpoint of its recommended dose at 10 weeks with the use of both fixed effects (29) and random effects (30) models. Ten weeks was chosen as the time point because this value required no extrapolation beyond the observed data for any drug.

Finally, we used pairwise comparisons for the estimated weight changes at 10 weeks at the standard dose of each compound. The significance of differences was tested with a z statistic. The quantity (θ_{i –} θ_j)/(SE²[θ_i] + SE²[θ_j]) is asymptotically (in the number of subjects not the number of means) distributed as a standard normal deviate, where θ_i and θ_j are the estimates of weight change for the ith and jth compounds, respectively (29). To account for multiple comparisons, we used Monte Carlo simulation with 100,000 simulated data sets to determine the z value that, given the number of tests being conducted, would hold the overall alpha rate to the two-tailed 0.05 level. The simulated data were generated from a model with normal distribution based on the sample sizes we had. (For the concept behind this approach, see reference 31.) The critical z value obtained was 3.31. Therefore, any pairwise comparison yielding a z statistic greater in absolute value than 3.31 is statistically significant even after accounting for conducting multiple comparisons. This is slightly less conservative than the 3.41 required for the ordinary Bonferroni correction.

This study has several limitations. First, standard errors were calculated under the assumption that all observations were independent, which was not true in every case because some studies assessed subjects at repeated time points. When such data were available, we included all data points in the interest of using all available information. Our estimates of weight change with use of the weighted least squares method remain accurate (i.e., unbiased), but their standard errors may be too small. Ordinarily, one would take this dependency into account through the use of generalized least squares estimation (26). Unfortunately, generalized least squares implementation requires knowledge of the covariance structure among the observations, and this information was not available. Therefore, the standard errors presented here and the significance levels based on them may, in some cases, be biased. To estimate the plausible degree of this bias, we assumed that all dependent observations had a correlation as high as 0.90 and conducted the fixed effects regression analyses through generalized least squares. The largest putative change in standard error for any drug occurred with chlorpromazine and was 59%. For no other drug did the increase in estimated standard error exceed 4%. Thus, this sensitivity analysis suggests that our standard errors are unlikely to have been under­estimated to any substantial degree.

A second limitation concerns our inability to examine the extent to which weight change with antipsychotic drugs varied as a function of patients’ characteristics, such as age, sex, and initial body mass index. Unfortunately, the limited information presented in each study on the distributions of age, sex, and starting body mass index and the limited number of studies available for each drug precluded inclusion of terms for such patient characteristics in the metaregressions.

A third limitation is that for most drugs, insufficient information was available to provide precise estimates of weight change when patients were on the drug for extended periods of time, such as 6 months or more. Although we initially attempted to calculate such estimates, this frequently required extrapolations outside the observed range of data, and the resulting estimates were often extremely imprecise (i.e., had very large standard errors).

To our knowledge, this is the most comprehensive literature synthesis on antipsychotic-induced weight gain to date. Although it is plausible that some studies assessing the effect of antipsychotic medications on body weight were not discovered by our literature search, our procedures kept this to a minimum. We conducted a thorough search of the electronic literature and made efforts to access undetected literature through the “Invisible College” and formal contacts with pharmaceutical companies. Moreover, we conducted electronic searches of databases that also include unpublished literature, such as Dissertation Abstracts International and PsychINFO.

A common problem in meta-analysis is inadequate and incomplete reporting of key information in the primary articles. We attempted to minimize the impact of such incomplete reporting by contacting authors when feasible. Moreover, for studies that did not yield sufficient information to be included in the formal quantitative synthesis, we still attempted to extract whatever information was available in these reports and provide that in a structured verbal overview (Table 2).

Our literature retrieval procedures also maximized the chances of obtaining relevant unpublished data. Publication bias is a commonly acknowledged problem in applied research. This problem confronts all literature reviewers, whether they conduct formal meta-analyses or not. We believe our efforts were quite strong in this regard and therefore should serve to minimize publication bias.

In the early days of chlorpromazine pharmaco­therapy, Planansky and Heilizer (61) reported that weight gain was associated with symptom improvement, and weight loss was associated with symptom deterioration. The subsequent availability of multiple antipsychotic medications has led to the observation that weight gain is a common side effect of antipsychotic treatment. There is some conjecture that the drugs which cause the most weight gain are the most effective. However, results are inconsistent and equivocal (58, 62–65), and further research is needed on this question.

How clinically meaningful are these degrees of weight gain? Some compounds were estimated to produce close to a 5-kg weight gain at 10 weeks. Furthermore, the shorter-term controlled studies usually included data on all subjects, whereas long-term use was usually restricted to individuals showing a positive therapeutic response to the drug. If therapeutic response and weight gain are correlated (which may or may not be true), then this would imply that the 10-week weight gain may be higher than we have estimated. On the basis of the compounds for which longer-term data were available (chlorpromazine, clozapine, and olanzapine), it seems clear that although weight gain might reach a plateau after a certain period (e.g., for olanzapine, after 4–5 months), total weight gain will still be much larger after periods longer than 10 weeks. Thus, weight gains far in excess of 5 kg may be seen in patients on long-term therapy. However, even a weight gain of 5 kg will represent a weight gain of more than 5% of initial body weight for the majority of individuals. To place this in perspective, it is useful to consider that a number of authoritative bodies, such as the Institute of Medicine (66), have suggested that weight losses of as little as 5% in obese individuals can result in clinically meaningful reductions in morbidity and risk of early mortality. It might be plausible, then, to expect that increases in body weight of as much as 5% would result in corresponding increases in morbidity and risk of early mortality.

Although the literature assessing the effects of weight gain on “hard” end points reveals a complex set of relations and modifiers, certain general conclusions can be drawn. For end points such as mortality (67, 68), incidence of cancer (69, 70), cardiovascular disease (71, 72), and diabetes (73), when factors such as smoking are accounted for, it appears that weight gains of 5% or greater during the adult lifespan are associated with important increases in risks. This is especially true for individuals who are overweight to begin with. Finally, although results are clearly preliminary, emerging data suggest that the drugs causing weight gain (i.e., clozapine and olanzapine) may, perhaps as a result, also be causing type II diabetes (74–77). Clearly, then, antipsychotics can induce medically meaningful degrees of weight gain.

Weight gain induced by antipsychotic drugs may also discourage patients from reliably taking their medication. This would, in turn, increase the likelihood of relapse. Although we are aware of no data that would allow precise quantification of the impact of weight gain on compliance with medication, we have observed that a number of patients complain of weight gain and occasionally report it as a reason for noncompliance. On the other hand, in studies conducted with olanzapine, Tollefson et al. (78) and Beasley et al. (60) found that for acute trials and studies lasting a year, drug discontinuation attributed to weight gain was quite rare. For example, in a 6-week study, Tollefson et al. (78) found that “none of 1,336 olanzapine-treated patients discontinued early because of weight gain.”

Given this background, it is important to consider methods for minimizing the impact of weight gain induced by antipsychotic drugs. One approach might be to implement weight control procedures with schizophrenic individuals who are taking antipsychotic medications. Several such efforts have been made in the past, including pharmacologic (79, 80) and nonpharmacologic (81–85) approaches. In some cases, particularly with subjects in inpatient settings, results have been good. However, results with outpatients are less clear, and more research on this topic is needed. Some investigators have even begun to explore the potential of amantadine in pharmacologic treatment specifically for neuroleptic-induced weight gain (86–88), but this is not a generally accepted treatment at this time.

The use of pharmacologic treatments of obesity with this population may present challenges. With one exception (89), all pharmacologic agents for the treatment of obesity that are currently available or likely to be released in the very near future achieve their effects by increasing noradrenergic, dopaminergic, and/or serotonergic activity (90). In contrast, antipsychotic medications typically achieve much of their effect by decreasing dopaminergic and serotonergic activity. Therefore, the use of pharmacologic agents to treat obesity in individuals with schizophrenia may exacerbate their psychotic symptoms (91–95). Indeed, weight loss itself has even been reported to provoke psychotic symptoms in rare cases (96). Therefore, any use of centrally acting pharmacologic agents to treat obesity in this population should be undertaken with the utmost caution and, in our opinion, be preceded by well-controlled clinical trials to establish efficacy and safety.

Finally, the selection of the right compound for the right patient might minimize the impact of weight gain with antipsychotic medications. There are schizophrenic individuals who are sufficiently thin that weight gain would likely be harmless and perhaps even beneficial (8). In such cases, not all weight gain will necessarily represent fat. Studies also indicate that weight gain is highest in individuals with a low baseline body mass index (60). Although these patients are rare, for such patients there would be little reason to avoid the use of drugs that produce greater degrees of weight gain. However, for patients with an average body mass index or higher or patients with a history of obesity, clinicians may wish to consider using compounds associated with less weight gain. The estimates provided in table 4 may help clinicians make such choices. The preceding notwithstanding, we wish to emphasize that weight gain should never be a sole reason for choosing one antipsychotic drug over another. Both therapeutic efficiency and other factors such as dose-related extrapyramidal syndromes should also be considered in drug selection. For many individuals the degree of risk imposed by the weight gain from a drug will not outweigh the degree of benefit achieved by alleviation of schizophrenic symptoms. In the end, clinical choices must be made on a case-by-case basis, with careful consideration of issues of weight, therapeutic efficacy, and other relevant factors.