Severe bereavement stress during the prenatal and childhood periods and risk of psychosis in later life: population based cohort study

BMJ 2014; 348 doi: http://dx.doi.org/10.1136/bmj.f7679 (Published 21 January 2014)
Cite this as: BMJ 2014;348:f7679

Recent rapid responses

Rapid responses are electronic letters to the editor. They enable our users to debate issues raised in articles published on bmj.com. Although a selection of rapid responses will be included as edited readers' letters in the weekly print issue of the BMJ, their first appearance online means that they are published articles. If you need the url (web address) of an individual response, perhaps for citation purposes, simply click on the response headline and copy the url from the browser window.

Displaying 1-5 out of 5 published

In Response to Jenny Bryden

“Thank you for an interesting and thought-provoking study. I wonder if some of the results may be influenced by the rules of statistics as much as the specifics of any neurodevelopmental effects, however?
In this paper, the presence or absence of a statistically significant difference in the OR from 1 is used to determine whether a risk factor is relevant. The chance of finding a statistically significant difference is at least partly determined by the sample size(1). The low absolute numbers of sudden deaths and suicides are to be welcomed, but do mean that the sample size for individual subgroups are small when the paper seeks to tease out the effect of timing of bereavements. For example, there are only 46 people with non-affective psychosis whose mothers were bereaved while they were in utero (another 46 had mothers who were bereaved prior to their conception). In affective psychosis, there were 11 people whose mothers were bereaved while they were in utero (with 15 whose mothers were bereaved before their conception)."

Authors' Response:
We appreciate the point made the commentator regarding sample size.
Even though this is one of the largest studies of this exposure thus far, the number of offspring who were exposed to maternal bereavement in a particular period of gestation and who developed a psychosis is not large. Thus, the results for specific time periods of gestation need to be interpreted with this caution in mind. Nonetheless, we remain confident about our findings for the following reasons:
First, we would agree that, in general, larger samples are more likely to find statistically significant effects, due to increased statistical power. We also agree that the relatively small numbers of sudden deaths and suicides mean we were unable to assess effect of suicide as a cause of death antenatally. However, we had found no main effect of any death at any time prenatally as a cause of any psychosis. The findings for death antenatally in the ‘all psychosis any prenatal’ group were based on larger samples (exposed n=115) and produced marginal effect sizes (OR 1.17 (0.97 to 1.41) which were non-significant. For exposure during specific gestational periods, the numbers are smaller, but the Confidence Intervals are still fairly narrow, even when the sample sizes are small i.e. 11 or 46. This suggests we had sufficient power to be confident in our conclusions about the point estimates.
Secondly, the commentator speculates that a larger sample would have increased likelihood of detecting a significant effect of small effect size, i.e. we would be more likely to find an effect regardless of its clinical relevance. Whilst it is true that logistic regression can overestimate odds ratios in small samples (Nemes et al 2009), in this large population-based cohort design, we feel confident that this does not affect conclusions because: first, our crude population level estimates could be constructed without logistic regression and lead to the same results i.e. this potential for bias does not affect the crude odds ratio; secondly, that in the prenatal analysis (where the commentator argues that small numbers apply), we have already discussed these findings as being of marginal clinical significance, and any bias would lead to these having been over estimated, and so our conclusions still hold; thirdly, we feel that our population estimates are based on large enough samples that the statistical bias inherent in the modelling procedure does not apply here.
Finally, in such a population-based design, we collect all the information on all population members exposed and unexposed. This equates to very large numbers: total exposed n=321 249; total unexposed n=625 745; and we cannot collect more data. The Odds Ratio, therefore, is the observed Odds Ratio for the population even with relatively small numbers in some cells.

Bryden Comments:
"Because the post-natal periods are so much longer, there are more bereavements, so it’s much more likely that any genuine difference from the control group (no exposure) will reach statistical significance."

Authors' Response:
We would agree that the small effects of no statistical significance in the shorter prenatal periods may conceal a relevant finding ie be a false negative. Focussing on statistical significance, however, would be more important if the effect sizes themselves were larger and non-significant with wider confidence intervals. However, across the ORs presented, the CIs were narrow and effect sizes moderate or small. We think this means it is less likely that we are making inferences of no effect where there are hidden false negative effects.

Bryden comment:
"The same phenomenon may be responsible for some of the differences between affective and non-affective psychosis. While there are 11, 117 people in the registry exposed to a severe bereavement, 1323 of these people developed a non-affective and only 556 an affective psychosis. This means both that the study has a higher power to find any genuine results in the causation of non-affective psychosis and also that point estimates are more likely to take extreme values. This may also be behind the differences in the literature connecting pre-natal stress with more common conditions (ADHD or depression) versus relatively rare ones such as psychosis."

Authors' Response:
There were, in fact, 11 117 exposed to death by suicide. Total exposed to death were n=321 249.
We are using a whole population and although psychosis is a fairly low prevalence outcome, increasing the size of a sample may increase the number of cases, but it would not affect the prevalence rate in the population and similarly, as we have used the whole population as the sample, we cannot double the sample nor increase the number of cases or the estimates.

Bryden comment
"In logistic regression, a logarithmic transformation of the odds of the outcome is used to construct the model(2). When you control for multiple confounders (at least 16 categories of possible values here, depending on how exactly the model was constructed) then the referent group becomes those who have null or referent values for all of these(3). Here, these would be people whose parents were both aged 25-29, Swedish and with 10-12 years of education, who had no siblings, no family history of mental illness and no parents receiving welfare payments. As more confounders are added, this referent group becomes smaller, the point estimates more extreme and the confidence intervals wider (especially when the outcome is relatively rare, as with psychosis)."

Authors' Response
The Swedish national registers represent an example of a high quality administrative database inasmuch as they have very few missing data on many key potential confounders. For example, there are almost complete data for many relevant characteristics, as mentioned in the paper. We acknowledge that in a logistic regression with multiple covariates, the comparison becomes the difference in outcome between exposure or not in the reference group of all the covariates combined e.g. male and in the lowest age group and… etc. and that multiple confounders in a model, in general, reduce the size of the comparator samples. However, as we state in the methods, we performed adjustments one by one for potential confounders initially. None of these made a material difference to point estimates except for family psychiatric history. Therefore, we only presented one column of adjusted analyses for practical reasons. In order to meet the definition of a confounder, these variables would need to be associated with the exposure, and for many of these variables this might seem implausible – for example, there is no reason to think that sex of the child is associated with our exposure of bereavement stress – and this might explain why adjusting for these variables did not lead to major differences between the crude and the adjusted odds ratios.
The fact that these problems can be present even in a registry study does not diminish its quality but simply underlines the difficulty in researching the rare outcomes of rare exposures."

We agree wholeheartedly with this statement.

Nemes et al Bias in odds ratios by logistic regression modelling and sample size. BMC Medical Research Methodology 2009, 9: 56

1. Calculation of Required Sample Size In Kirkwood B and Sterne J; Essential medical statistics (2nd Ed); Oxford; Blackwell Science; 2003
2. LaMorte W; Regression analysis: controlling for confounding; Boston University School of Public Health. Available from http://sphweb.bumc.bu.edu/otlt/MPH-Modules/EP/EP713_Regression/EP713_Reg...
Accessed 28/1/14
3. Logistic Regression: Controlling for confounding and other extensions In Kirkwood B and Sterne J; Essential medical statistics (2nd Ed); Oxford; Blackwell Science; 2003

Rollo Sheldon's comments:
“I am surprised that a significant confounder was not mentioned in Abel et al.'s research - that of childhood abuse.
Varese et al. (2012), in their Meta-analysis showed a significant contribution of childhood emotional, physical and sexual abuse towards later psychosis, but not parental death. Their population attributable risk was 33% (16-47%).
Perhaps this is related to epigenetics (i.e. methylation of the glucocorticoid receptor gene NR3C1- (Perroud et al 2014)), a wonderful fusion of the biological and the psycho-social?”

Authors' Response
The commentator is right to highlight how there is an increasingly recognised link between childhood adversity (in the broadest sense) and later psychosis risk. The Varese et al paper is a meta-synthesis including a range of population and non-population based studies. It does not directly assess whether or not childhood adversity per se is a potential confounder of the association between antenatal or postnatal bereavement stress and later psychosis and there is no particular reason to assume that childhood adversity per se is likely to increase risk of the exposure i.e. loss of a close family member. Of course, childhood adversity related to bullying, social deprivation etc consequent on close family bereavement may, as we suggest, be in part mediating the association we describe.

References:
Varese et al.,Childhood Adversities Increase the Risk of Psychosis: A Meta-analysis of Patient-Control, Prospective- and Cross-sectional Cohort Studies. Schizophrenia Bulletin (2012) doi: 10.1093/schbul/sbs050 First published online: March 29, 2012

Nader Perroud et al.,Childhood maltreatment and methylation of the glucocorticoid receptor gene NR3C1 in bipolar disorder. British Journal of Psychiatry January 2014 204:30-35; doi:10.1192/bjp.bp.112.120055

Competing interests: None declared

KM Abel, Professor

Hein Heuvelman, Lena Jorgensen, Susanne Wicks, Ezra Susser, Johan Hallqvist, cecilia Magnusson, Chrsitina Dalman

Universirty of Manchester, Centre for Women's Mental Health

Click to like:

Dear Editors,

Verbal, physical, psychological, sexual abuse, are even more stressful to women.

They are also widespread. [1][2]

Psychosis in children of these women could be even more prominent and persistent.

References

[1] http://www.bmj.com/content/346/bmj.f4077
[2] http://www.bmj.com/content/346/bmj.f4077?tab=responses

Competing interests: None declared

Stavros Saripanidis, Consultant in Obstetrics and Gynaecology

Private Surgery, Thessaloniki, Greece

Click to like:

Thank you for an interesting and thought-provoking study. I wonder if some of the results may be influenced by the rules of statistics as much as the specifics of any neurodevelopmental effects, however?

In this paper, the presence or absence of a statistically significant difference in the OR from 1 is used to determine whether a risk factor is relevant. The chance of finding a statistically significant difference is at least partly determined by the sample size(1). The low absolute numbers of sudden deaths and suicides are to be welcomed, but do mean that the sample size for individual subgroups are small when the paper seeks to tease out the effect of timing of bereavements. For example, there are only 46 people with non-affective psychosis whose mothers were bereaved while they were in utero (another 46 had mothers who were bereaved prior to their conception). In affective psychosis, there were 11 people whose mothers were bereaved while they were in utero (with 15 whose mothers were bereaved before their conception). Because the post-natal periods are so much longer, there are more bereavements, so it’s much more likely that any genuine difference from the control group (no exposure) will reach statistical significance.

The same phenomenon may be responsible for some of the differences between affective and non-affective psychosis. While there are 11, 117 people in the registry exposed to a severe bereavement, 1323 of these people developed a non-affective and only 556 an affective psychosis. This means both that the study has a higher power to find any genuine results in the causation of non-affective psychosis and also that point estimates are more likely to take extreme values. This may also be behind the differences in the literature connecting pre-natal stress with more common conditions (ADHD or depression) versus relatively rare ones such as psychosis.

In logistic regression, a logarithmic transformation of the odds of the outcome is used to construct the model(2). When you control for multiple confounders (at least 16 categories of possible values here, depending on how exactly the model was constructed) then the referent group becomes those who have null or referent values for all of these(3). Here, these would be people whose parents were both aged 25-29, Swedish and with 10-12 years of education, who had no siblings, no family history of mental illness and no parents receiving welfare payments. As more confounders are added, this referent group becomes smaller, the point estimates more extreme and the confidence intervals wider (especially when the outcome is relatively rare, as with psychosis).

The fact that these problems can be present even in a registry study does not diminish its quality but simply underlines the difficulty in researching the rare outcomes of rare exposures.

1. Calculation of Required Sample Size In Kirkwood B and Sterne J; Essential medical statistics (2nd Ed); Oxford; Blackwell Science; 2003

2. LaMorte W; Regression analysis: controlling for confounding; Boston University School of Public Health. Available from http://sphweb.bumc.bu.edu/otlt/MPH-Modules/EP/EP713_Regression/EP713_Reg...
Accessed 28/1/14

3. Logistic Regression: Controlling for confounding and other extensions In Kirkwood B and Sterne J; Essential medical statistics (2nd Ed); Oxford; Blackwell Science; 2003

Competing interests: None declared

Jenny R Bryden, CT3 in Psychiatry

Royal Cornhill Hospital, Cornhill Road, Aberdeen AB25 2ZH

Click to like:

I am surprised that a significant confounder was not mentioned in Abel et al.'s research - that of childhood abuse.

Varese et al. (2012), in their Meta-analysis showed a significant contribution of childhood emotional, physical and sexual abuse towards later psychosis, but not parental death. Their population attributable risk was 33% (16-47%).

Perhaps this is related to epigenetics (i.e. methylation of the glucocorticoid receptor gene NR3C1- (Perroud et al 2014)), a wonderful fusion of the biological and the psycho-social?

References:

Varese et al.,
Childhood Adversities Increase the Risk of Psychosis: A Meta-analysis of Patient-Control, Prospective- and Cross-sectional Cohort Studies
Schizophrenia Bulletin (2012) doi: 10.1093/schbul/sbs050 First published online: March 29, 2012

Nader Perroud et al.,
Childhood maltreatment and methylation of the glucocorticoid receptor gene NR3C1 in bipolar disorder
British Journal of Psychiatry January 2014 204:30-35; doi:10.1192/bjp.bp.112.120055

Competing interests: None declared

Rollo J G Sheldon, CT2 Psychiatry

Sussex Partnership NHS Foundation Trust, Swandean, Arundel Road, Worthing BN13 3EP

Click to like:

Loss of a dear one at a tender age I could perceive in the eyes of one of my nephews whose father was about to be buried. As the body was in the process of being buried I saw the young boy gazing at the coffin and the sky giving a perplexed look. He knew that his father is dead. He could not understand the concept of death and dying. Yet that look was more expressive than the volume of words we could write or communicate. Science can provide data. Only the experience of children who have undergone such losses know what the death of a parent means.

The environment and the love of those who bring such children do play a role in getting them out of such shackles of shock and bereavement. Such a mental agony abates with time and the human resilience fights such tragedies. Life and death are part and parcel of life. We know many children have been taken away from parents and sold in the market for begging. The children are mutilated or amputated or blinded to beg on the streets. These children grow up to become emotionally paralyzed and with no sympathy towards humans and values of human life. The tragedy of natural bereavement pales into significance before the early emotional killing of these children who are the victims of human trafficking. Science does not attach value to scientific truths. Science cannot explain the emotional tragedy of these children who lose their parents.

Competing interests: None declared

Dhastagir Sheriff, Professor

Faculty of Medicine, Benghazi University, Benghazi

Click to like:

THIS WEEK'S POLL