Hi, Heber.

The systematic review and the meta analysis, is not the same thing, the first, grouping, evaluates and synthesizes systematically all relevant information available to answer a research question, with a protocol established previously, this with the objective that can be reproducible.

The meta-analysis is the statistical method to group the quantitative data that can be extracted derived from the systematic review, this using two main models: fixed and random effects.

The heterogeneity measure derived from the Q statistic, and the H, I2, T2 indices are obtained from the variance between the studies of the response variable, however, the decision to determine the heterogeneity should not be based only in the Values mentioned above, but also of the design of the studies and their response variables, the objective and the population of interest; therefore if studies show heterogeneity can apply random effects for the meta-analysis.

A problem in dentistry is the large number of measurement scales (ie you can not compare pears with apples) and the lack of complete data reporting, this greatly limits the possibility of making valid comparisons.

Finally to answer your question: the systematic review aims to gather all the empirical evidence that meets previously established eligibility criteria, in order to answer a specific research question. It uses systematic and explicit methods, which are chosen in order to minimize biases, thus providing more reliable results from which conclusions can be drawn and decisions made (Antman 1992, Oxman 1993)

The answer above gives sound guidance. However, I think the answer to your question can also be made on a different level. We often treat certainty / uncertainty as dichotomy whereas the reality is that science exists on a continuum of uncertainty - from a lot to a little. Just as non-systematic reviewers have often failed readers in the past by giving a false appearance of certainty by selective and uncritical reading of evidence that supports prior beliefs, sometimes systematic reviewers simply say there is uncertainty because there is but do little to attempt to make a more nuanced assessment - the degree to which uncertainty has been reduced by their review.

Of course, the reality is that there is a lot of uncertainty out there, much evidence is heterogeneous and does not support meta-analysis and many revues do indeed do absolutely nothing to resolve uncertainty.  Sometimes, although not always, this is because the review was pretty pointless in the first place - we knew there was insufficient evidence before we started. The authors of some of these reviews may be being a little 'lazy' for want of a better word, but at least they are being honest!

I have learned to ignore the conclusions of metaanalysis because in obstetrics, the conclusions often do not reflect the research findings but rather whatever the author(s) wanted to find in the first place.  

This is true for reviews of low risk birth outcomes at homebirths, uterine ruptures after cesarean, causes of cord prolapse .  The authors often ignore obvious and relevant data like  BMI or birth weight, or use of induction.   Just bad science.  However, 

Regardless of the conclusions, the authors have done a big job of collecting relevant research and i like to read the data and findings, and draw my own conclusions.

Hi Heber,

Just a quick point about interpreting heterogeneity - You also need to initially consider the graphical depictions of the included studies' confidence intervals for the amount of overlap present. You would expect statistical heterogeneity (the I2 statistic as an example from your question) to be present if there is poor overlap of the confidence intervals. This then helps you interpret the formal statistical exploration of heterogeneity that the authors have completed.

To answer your question, studies need to be suitably homogenous, in terms of clinical diversity to provide a meaningful summary. If there is a lack of consensus in terms of measurement scales for a particular outcome measure, this will be difficult. If, however, there is agreement in outcome measure (such as a simple binary yes/no, presence/absence) then meta analysis is more likely to be possible as there is less room for variance across studies (e.g. you either have a dental carie or not). The following may be of interest to you. The authors completed meta analysis for one of their objectives and had a very low I2. Note the binary outcome measure: http://onlinelibrary.wiley.com/doi/10.1002/14651858.CD005296.pub3/full

In this way, heterogeneity is also linked to the review question and outcome measures. 

Best wishes,

Agi 

Heterogeneity is the between-study variation and is often calculated as the I2statistic measuring the proportion of variation in the combined estimates due to between-study variance (i.e. inconsistency). An I2 value of 0% indicates that no inconsistency (variation beyond chance) was seen between the results of individual trials, and an I2 value of 100% indicates maximal inconsistency. Furthermore the between study variance can be expressed as tau-square measuring inconsistency at the same scale as the effect size and standard error.

Prediction intervals encompassing the effect of a future study with 95% certainty can be estimated in order to evaluate the quantitative impact of inconsistency according to the method of Higgins et al using the between-study variance and the standard error.

Meta-regression analysis aims to relate the treatment effects recorded in different trials to the characteristics of patients (e.g. age, BMI), interventions (e.g. intensity, duration), study design (e.g. individual randomization or cluster randomization), comparison group (active or passive) or outcome measure in order to explain heterogeneity. This means that we can investigate the impact of study covariates. These meta-regression analyses can then provide estimates for groups with different disease severity, intervention, intensity and dose and these meta-regression analyses can provide evidence-based recommendations. A relevant study-level covariate can then be defined as one able to decrease inconsistency  measured as the I2 statistic and thus the between-study variance (tau-square), (meaning including this covariate in the analysis are explaining part of the between study-variance  and an estimate with lower heterogeniety can be found).

Best Carsten

agreed with Peter