Dear all,
My question is about the appropriateness of using metareg for a dose response meta regression given the limitation that glst requires further data that exclude half of my studies.
Im carrying out a meta-regression as part of a systematic review looking at the association between level of area deprivation (measured using a varying (3 to 5) number of categories) and place of death (home v institution). I want to determine whether there is a linear dose response association between deprivation and place of death - the hypothesis is that hgiher levels of deprivation is associated with higher odds of death in institution.
I have 17 studies with 66 categories with a corresponding dose and adjusted odds ratio (excluding the reference category)
Im using 2 commands both from Stata Journal:
metareg
glst
For info (but not important to the question) - im assigning a dose to each deprivation category using an established method that applies a mean relative rank - so for example if category 1 contains 30% of the sample, and category 2 contains a further 34% of the sample, the dose for cat 1 will be 15 (30/2) and the dose for cat 2 will be 32 (30+34/2).
I understand that glst is the most appropriate method to use for a dose response meta-regression because it is carries out a 2 stage model (either fixed or random effects), first estimating the within study coefficient and then pooling these to give the overall coefficient.
Alternatively metareg estimates a random effects model of the coefficient, assuming each data point is independent - an assumption that is of course violated because each study contributes multiple (between 3 and 5) data points/odds ratios.
So glst seems the more statistically robust command for use in dose response meta-regression, but it comes at a cost! glst requires additional information to estimate the within study variance - in this analysis data on the (unadjusted) number of cases and non-cases in each deprivation category is required. Less than half of my studies contain this additional data. I can contact authors to try and source this data but it is unlikely that data will be returned for all studies.
A further point is that some of my studies use the most deprived group as the reference and some use the least deprived.
So i carried out some cursory sensitivity analysis comparing results from metareg and glst on the following models
/*glst - all studies (only 8 with complete data)*/
glst logor dosec, se(selogor) cov(n cases) pfirst(id study) eform
/*metareg - all 16 studies together*/
metareg logor dose, eform wsse(selogor) graph randomsize
/*glst - 4 studies with richest as ref */
/*glst - 4 studies with poorest as ref*/
/*metareg - 4 studies with richest as ref*/
/*metareg - 4 studies with poorest as ref*/
They all give the same pooled estimate and confidence intervals - not exactly the same but when rounded to 2 decimal points they are the same.
The other advantage of metareg is the output includes a 'bubble plot' (attached) which is a useful summary of all the data points.
My questions (finally!):
1. Is it appropriate for me to use metareg on all of the 16 studies together (mixing those that use the poorest group as ref with those that use the highest group as ref) and report details of the sensitivity analysis alongside? The graphical display from this is particularly useful and nice graphic.
2. Is there any way for me to account for the dependence between the data points using metareg? I think not, but perhaps i am missing something?
3. I can see that combining studies with different ref groups is potentially problematic for metareg , but because glst estimates the within study trend first - is glst able to combine studies using different reference groups?
I realise there are other things to consider with this analysis including how well the models fit the data and whether quadratic terms might be more appropriate.
I should also say that this dose response analysis is a secondary or tertiary aim for the systematic review - so it would be nice to summarise the data graphically and with a pooled trend estimate but this is really only descriptive in purpose and is obviously limited by the fact that this is observational messy data. It interesting/useful because although a lot of data has been collected on this topic, no one has summarised it before, and although not surprising that there is a social gradient at the end of life just as there is in almost every other area of health there is very little attention paid to this source of inequality within end of life care planning and provision.
Apologies for the length of this post. I have searched extensively for answers to these questions and although there is a great deal on how to use metareg and glst and many examples i have found nothing that addresses the nuances of these questions.
I would be extremely grateful for any input.
Many thanks,
Joanna
My question is about the appropriateness of using metareg for a dose response meta regression given the limitation that glst requires further data that exclude half of my studies.
Im carrying out a meta-regression as part of a systematic review looking at the association between level of area deprivation (measured using a varying (3 to 5) number of categories) and place of death (home v institution). I want to determine whether there is a linear dose response association between deprivation and place of death - the hypothesis is that hgiher levels of deprivation is associated with higher odds of death in institution.
I have 17 studies with 66 categories with a corresponding dose and adjusted odds ratio (excluding the reference category)
Im using 2 commands both from Stata Journal:
metareg
glst
For info (but not important to the question) - im assigning a dose to each deprivation category using an established method that applies a mean relative rank - so for example if category 1 contains 30% of the sample, and category 2 contains a further 34% of the sample, the dose for cat 1 will be 15 (30/2) and the dose for cat 2 will be 32 (30+34/2).
I understand that glst is the most appropriate method to use for a dose response meta-regression because it is carries out a 2 stage model (either fixed or random effects), first estimating the within study coefficient and then pooling these to give the overall coefficient.
Alternatively metareg estimates a random effects model of the coefficient, assuming each data point is independent - an assumption that is of course violated because each study contributes multiple (between 3 and 5) data points/odds ratios.
So glst seems the more statistically robust command for use in dose response meta-regression, but it comes at a cost! glst requires additional information to estimate the within study variance - in this analysis data on the (unadjusted) number of cases and non-cases in each deprivation category is required. Less than half of my studies contain this additional data. I can contact authors to try and source this data but it is unlikely that data will be returned for all studies.
A further point is that some of my studies use the most deprived group as the reference and some use the least deprived.
So i carried out some cursory sensitivity analysis comparing results from metareg and glst on the following models
/*glst - all studies (only 8 with complete data)*/
glst logor dosec, se(selogor) cov(n cases) pfirst(id study) eform
/*metareg - all 16 studies together*/
metareg logor dose, eform wsse(selogor) graph randomsize
/*glst - 4 studies with richest as ref */
/*glst - 4 studies with poorest as ref*/
/*metareg - 4 studies with richest as ref*/
/*metareg - 4 studies with poorest as ref*/
They all give the same pooled estimate and confidence intervals - not exactly the same but when rounded to 2 decimal points they are the same.
The other advantage of metareg is the output includes a 'bubble plot' (attached) which is a useful summary of all the data points.
My questions (finally!):
1. Is it appropriate for me to use metareg on all of the 16 studies together (mixing those that use the poorest group as ref with those that use the highest group as ref) and report details of the sensitivity analysis alongside? The graphical display from this is particularly useful and nice graphic.
2. Is there any way for me to account for the dependence between the data points using metareg? I think not, but perhaps i am missing something?
3. I can see that combining studies with different ref groups is potentially problematic for metareg , but because glst estimates the within study trend first - is glst able to combine studies using different reference groups?
I realise there are other things to consider with this analysis including how well the models fit the data and whether quadratic terms might be more appropriate.
I should also say that this dose response analysis is a secondary or tertiary aim for the systematic review - so it would be nice to summarise the data graphically and with a pooled trend estimate but this is really only descriptive in purpose and is obviously limited by the fact that this is observational messy data. It interesting/useful because although a lot of data has been collected on this topic, no one has summarised it before, and although not surprising that there is a social gradient at the end of life just as there is in almost every other area of health there is very little attention paid to this source of inequality within end of life care planning and provision.
Apologies for the length of this post. I have searched extensively for answers to these questions and although there is a great deal on how to use metareg and glst and many examples i have found nothing that addresses the nuances of these questions.
I would be extremely grateful for any input.
Many thanks,
Joanna
Attached Files
Comment