Health Studies User Conference 2027: Parallel session B

But welcome back for the afternoon of the um health studies users conference today. Um we've got five talks to go through in this parallel session and we're very um excited to have um so much interest in presenting at this conference. So thank you to all the speakers. Our first speaker is um Kristoff um Hen from University of Oxford and the European Commission uh joint research center. He's a post-docctoral researcher on the program for work welfare reform and mental health in the department of global health and social medicine as well as at um the ESRC center for um society and mental health. He's a quantitive social scientist working at the intersection of health and social policy and he I'll leave pass over to him to introduce his talk and take it away. Thanks so much Patty and uh thanks for everybody organizing this conference uh despite the heat and uh yes um so the work I represent today is from my postto at King College London I am now um I changed job I'm now at the European Commission actually in Brussels as a researcher at the joint research center um of the European Commission and I'm continuing my work that I'm presenting today in a different institution and just to say that my my research that I present today is is unrelated to my current role in this space for my um my role at King's College. Uh let's get started and I try to share my screen. Um share. All right. Okay. I hope you can see this. Um the paper I will present today is a working paper. I'm working on this closely together with with Professor Ben Geger. I think Ben, you're also in the in the audience, which uh of course it makes me more nervous, but okay. Um yes, and uh we took a take a closer look at um mental health trends in in England for the for the last 20 years. Um and uh what we do in this research um is to kind of ask the question when it comes to mental health we have we have this uh debate about mental health crisis and and really related to you know young people's mental health in in England UK but you know all over Europe as well but um we sometimes speak about different aspects when we speak about you know is mental health in crisis is mental health getting worse um but there actually different elements different indicators we can we focus on and and this is rarely done. Um because on the one hand we can look at actually distress or um prevalence rates of depression, anxiety, how do they develop? But there's more to this than just measuring population level um population level mental distress. But there's also the question whether people selfidentify as having a mental illness and this aspect of medicalization. So do people interpret that distress as a mental illness? And also uh the third one we look at is exclusion. So are people with mental health problems then also excluded and limited in uh in the labor in the work place and and and otherwise and there's of course still also more variables but we focus on these three and we asked the question trends how are these three different indicators and different elements of mental health developing over time and what can we learn um and previous research um has um and please patty also do interrupt me I'm not sure if I I think I have 10 15 minutes to speak Um >> sorry you have 18. >> Uh this is and there's questions uh >> two minutes questions. Yeah >> afterwards. Okay. Sorry. Sorry about this. I should have Yes. So um when it comes to these different these three different elements of dimensions of mental health. So distress uh mental distress we do have evidence prior to to our study that there is a rise of mental distress um and also anxiety depression um particularly um in the 2010s in the times of austerity um so there's evidence on this there's also some evidence on um on this question of of self-reported long-standing illness this element of medicalization do people then say or interpret their distress as a mental illness We do have some evidence that this is increasing but maybe a bit less clear and on the whether this then translate into activity limitations. This we have you know not much at all in England and we know from some other countries um uh you know USA, Canada and others that we see that these trends may actually diverge. it might be that the stress uh measure goes up and this element of selfidentified illness you know goes up even further or in other countries diverges from from the other trends. So what we ask in this research is simultaneously looking at these three uh indicators um how have they developed in the past 20 years and what can we learn for for policy and further research. Um our uh data source is the health survey for England. Many of you will be familiar with this. Um, uh, we are focusing on the waves from 2003 until 2022. 2022 was recently released. We're still waiting for another wave in 2024, so this might get a bit extended. Um, and the HSSE is a nationally representative English survey with repeated cross-sections. So it's not longitudinal but it's repeated cross-sections of really you know large sample each time the it's taken and it's you know Ben would correct me there but but um often taken or used to be every year during co it's it was a bit more delayed but it's really really really informative data source for anything on prevalences and trends over time and it's an excellent source for this and we focus here on the let's say working age population from 18 to uh 64 before I mentioned already uh sorry to repeat myself the three measures we're interested in the psychological distress we use the general health questionnaire it's a GHQ which has a where people are scored from based on 12 different questions um each of them gives people a potentially a score for mental distress such as worrying or problem sleeping depression some well-being questions also but in the end of the day the people get a a score the higher score means more distress and a threshold of four or higher in this study we use as a caselessness. So a case of high distress of course we do some robustness checks for this as well but when we here speak about cases it's going to be four or higher. There's a question then on long-standing illnesses. So people are asked in the survey do you have a long-standing illness um and what type of illness is that? And we look at the the mental health uh classification here um for the sort of medicalization and selfidentified mental illnesses and for people who say I do have a long-standing mental illness are then also asked about do you have activity limitations due to this a lot or a little um and the three questions how have these three trends developed uh over time uh for these 20 years have they developed differently for different age groups do these trends trends differ if we control for social demographic factors like gender, age, we also tried education um in another analysis. So do they hold adjusting for other factors and among the people who have high distress does the inner identification change for for this among this this group and u I'm not going to go much into the sample it's pretty representative uh um and there's rates to adjust for for survey dropouts and so on so it's it's pretty uh pretty accurate um 109,000 observations for all these 20 years straight into some of the findings and I apologies that this this slide is is bit full but the main one of the the trends we look at is really the top left you see so the one you see that says overall population trends I walk you through briefly we have the black the black line this is the caseness of GHQ the cases of psychological distress the percentage of people in the population who who report four or higher on the GHQ scale and we see that uh that it does from left to right 2003 to 2022 on the right that it does increase uh and does increase from 13.6 to 19.1% over these 20 years. We then focus on the on the uh on the blue line which is this selfidentified mental long-standing mental illness. Um uh the indicator has changed uh somewhat. we've been pointed to by another researcher is the the question has changed in the slight bit in 2012 I believe and so we changed this analysis so um to reflect that they are not exactly the same question but they are very similar um that's why we have a you know full line and the dotted blue line and this one again we see or we do see it it also goes up it goes up quite strongly more than threefold um so from 3.7% of the general population saying I have a long sitting illness to 12% in 2022. Um the blue the red dotted line uh on the top left panel panel A we see um uh also do you see the the the dotted line the lighter red is the a little limitations due to mental uh illness. Uh and this goes up slightly as well whereas the a lot limitations are a bit more flat. Um though of course this indicator is also a bit conditional on the other questions. So it might be also that we we have definitely some limitations here how we measure this. Uh we then in panel A, B, C and D look at um more closely into the age groups um where we see um uh again GHQ the black the panel B for different age groups the young middle age and older workers uh also we see this increase and particularly on the panel C the bottom left this long-standing illness identification people identifying as having long sing this is really the youngest group increases a bit more strongly than than the other groups. Um we then also formally test this. Um this now is a uh we use regression models to to yeah control for in this case for for age and educate um age and and gender or sex. Um I think bi biological sex I believe is recorded that's why anyways but uh we also got for education sensitivity analysis. Um here you see the each of these we see again the same indicators caseness long-standing limitation long-standing illness and limitations from the baseline in this case the baseline is 2013 because the longstanding the the limitations variable only is collected from 2013 so that's why also the what I showed you before is not exactly the same time horizon for all of them but here we have it from 2013 and again the results from the regression analysis um uh show if you look at the top pane panel saying like overall uh we see that uh caseness increased by 3.7 percentage points. So from from a from 15% 6% plus 7 uh 3.7 whereas long-standing illness increased by five uh.7 percentage points and so on. So we again see this increase but in in in the stress but the long-standing illness identification it just increases more more strongly uh overall in terms of the size um of changes that we see and this is um yeah uh also pretty much the main message of one of the main messages of our findings. Uh finally looking the final analysis I will show you then then you've survived all the different charts that that we have here. uh looking at people because it's of course interesting to see okay um are we talking about what is really happening here how are these different trends relating to each other um and we uh look at here we look at people among the people who have uh GSQ cases and we do some sensitivity also to look at different thresholds so among this group is there a higher um identification of long mental illness so kind of like uh looking at people with distress is this group more likely to to say they have a long-standing illness and yes we do see this this trend um trend here going up the self identification of long-standing illness increases among the people with high distress whereas the the the limitations they also increase the a little limitations also increases um but not the severe limitations um what does it all mean and of course I'm really really curious this is a a working paper now but it's very fresh results So we very curious about what you you say about this some let's say interpretations and and and findings um um we do see that the stress is increasing it increases interestingly it's not a co story it's a story about biggest increase in the stress happens uh before co um and it does happen to some extent to more to young people but it's really in the whole population the phenomenon um we see in the exclusion uh limit activity limitations. It does also increase particularly this this mild limitations people report due to mental illness and we um this could also be you know related to you know the way um the stress is experienced in workplaces in everyday life how people might also be excluded more and more um and we and related to job quality changes. So we will do some discussions also here um to interpret this trend although of course it's not not a cause analysis and we see rising in uh strong increase and this identification of long-standing having a long-standing illness and this rose very strongly almost three-fold and much stronger than than the distress by itself but we do want to stress that both is happening in our data we see there's both an increase in distress mental distress and also an increase in this limitation in this um boxing in this identification. We see both. They're not mutually exclusive. They're both happening. And this is an important message here that they can it's not either or. Um uh and finally in terms of what is actually happening that medic long-standing illness is increasing. Um different things different elements could drive this. We don't we cannot sort of test those but those are certainly also future research to to look at. um you know for sure improved recognition and treatment. you know this greater awareness stigma and these things could could contribute uh in a also in a way to to more people um reporting this um but also this element of culture medicalicalization um where um you know we have we it could also be due to somewhat changing languages uh around mental mental illness and um that you know the interpretation of mental distresses mental illness is al is increasing by itself And finally uh what we call assistance medicalizations that people in uh in the welfare state like in in people increasingly uh who are who who don't feel well or have mental health problems increasingly have to have this mental illness label to get any kind of support to get um welfare support and then would be much more um uh sort of elaborate to to to discuss this but we do see that people are excluded from welfare systems if they don't um declare a mental illness or medical condition. So this could also drive people to increasingly report a mental illness because they have to because otherwise they don't get get support that they need. Um and of course f future research is needed to unpack these things. And I think it hopefully offers a basis for future research and and discussion for both policy and um and of course research. Right? I mean I think I gave you this already the conclusion. There are several limitations. You know it's not a cause analysis. Uh we don't have clinical information on diagnosis. These are all self-reports. Um and and yes and it's related only to England. Um, but I think I leave it here and I look forward to your your comments and also I have my email address and Ben's email address for f future you know discussions as well. Thank you. >> That's that's great. Thanks so much Kristoff. That was um really interesting presentation and fantastic timing. So we've got time still for um a couple of questions for Kristoff. If anyone um would like to ask anything um please put them in the question and answers box um just give you a moment to do that. If there's anything that you would um like to ask um if it doesn't come to you straight away. Oh, I've got a question here already. >> Understandable sometimes when you speak online you don't know if people can follow but I hope I hope you did. No, there's a question there um asking, sorry if they missed it, but did you compare trends before and after CO? >> Um yes. So um I mean we did not statistically that's a good part. We did not statistically test um uh but we cover the whole period. So that's why uh why we do because we we cover the whole time period. So yes, but we could still formally test the the pre postco versus before on the sort of the effect sizes of data we see. It's quite clear um and I think every statistical test would also show this that it's before co the stronger increase happens but we could it's a good point we should do a formal statistical test also for just these periods but then if that's what you what you mean here yeah >> there's also a interesting question um here from tiara asking how is longstanding um defined in the survey and do you have an idea of or if there is data available in the survey to see if um how these trends differ by type of mental illness or whether self-reported activity limitations may be influenced by the receipt or lack thereof of mental health services. >> Yeah. Um I have to go back for the for the precise survey item. I would have to go back um to to look this up. Um but yeah it's an open question on do you have any long-standing lungs illness and it's open in terms of people also report yes and then they have an ability to report other um also physical health conditions right so it's not exclusive to mental health and I think we cannot uh mental health I think one category one or two so we cannot look at different types um I think um but yeah um and it it it could it would be interesting to link it to um to mental health mental health services received. That's I think not a strong item that we could use, but a good point for me to double check this as well. >> Yeah. >> Um and there's a question there from Michael asking um anecdotally that that middle class girls are increasingly seeking CAM's help um you know through the social services or local support. So, how does um your demographic data or what does your demographic data suggest in regards to that? Was there any class indicators for women, young girls? >> Um uh we do um we do look at it separately right now. So, we look at we look at um men and women by gender and we look at age groups. um we don't so far really do an um I mean we use education as a as a as a proxy for so so economic status in a way but um would be interesting to look at the intersection here we we haven't done that so far we do see that young people um have an increase and we do see that the increase in long-standing illness identification goes across education groups and also across people who are currently working and not working so it's quite consistent but it would be interesting to look at this in more detail but we haven't >> yeah and there's one more question the question answers box but I'm going to have to leave that for now Kristoff but if perhaps you want to write your answer in >> sure that'd be great and just just to say that there is a working paper on this now and yeah Ben and I we would love to get your your feedback and other other points as well and yeah thank you so much for for >> brilliant thanks very much Kristoff um next we have um Katie Sarah Taylor who is a research fellow at UCL um within equalize which is the ESRC center for life course health and equity where her work explores placebased inequalities and health and she recently completed her PhD so congratulations Katie in social epidemiology with the sock B center for doctoral training in bioosocial research so I'll pass over to you Katie now to um tell us about what you've been working on thank you very much. Um, yes. Hi everyone. I'm Katie. I'm hoping that you can see my slides in here. Okay. Um, and I'm going to be presenting about yeah adolescent mental health area level deprivation and young adults categorized as not in education, employment or training or this term neat. Um, in the UK. So if we start by defining the term meat, um typically it's defined as young people who are aged between 16 to 24, sometimes 18 to 24 um who are categorized as not being in education, employment or training. I've linked um uh a briefing that our center equaliz has done about using the terminology of neat um that some of our colleagues in strafly did um because typically neat is used to um us as a noun. So we would normally say like meet young people but actually speaking to young people who are in the category of neat um find this quite stigmatizing. So we're trying to move the language more to yeah rather than young neat people meet young people more to young people in the neat category. So why why is this um why is neat important? Um so in the UK 1 million um 16 to 24 year olds or the equivalent of one in eight um are in the neat category and this has risen since 2019. And roughly a decade ago, the UK had comparable rates of meat um to the EU average um whereas now we have the second highest rates in the EU, second to Romania. And we've seen other European countries manage to um decrease their rates whilst ours have increased. So as a comparison, Netherlands currently have much lower rates at 4% compared to the UK at 15%. We also know that um being in the neat category can have a scarring effect. Some recent research from the center for longitudinal studies at UCL found that um young people in the neat category um had worse um employment and health outcomes in midlife. Um and this was worse um the longer that they spent in the neat category. So we know that there are these negative consequences as well. So it's not surprising that there has been this heightened policy focus on um young people who are in the neat category. And in January this year, there was a call for evidence from the Department for Work and Pensions and they were looking into getting evidence around the drivers of um young people becoming um being in this new category. And they released an interim report uh at the end of May, so just a few weeks ago. Um and in the autumn they're going to be releasing another report looking more at the solutions that we can um involve to reduce the rates of meat. And in this report one of the main conclusions was that health appears to be a central driver of these rising meat rates. So among um the young people who are in the neat category um in 2025 44% of them had reported work limiting health conditions which is up from 26% in 2015. And among um disabled young people who are in the neat category, those reporting mental health as a primary condition has risen from 24% in 2011 to 43% in 2025. whilst those reporting physical health as a primary condition has decreased from 74% to 32% in the same time frame. So what this suggests is that mental health seems to be playing quite an important role in these um rising mat rates as both the increases in mental health and um increases in meat are co-occurring. So if we do assume that mental health predicts whether someone is more or less likely to be in the neat category later on, what we're interested in is whether everyone might experience this the same way or not or whether there could be some structural factors that might make this relationship differ. So we're particularly interested in the place context um more specifically area deprivation. So what we're really interested in is whether this relationship between um specifically adolescent mental health and later neat risk varies by place context or area deprivation. So we had six research questions. The first being whether adolescent mental health shapes need risk. The second the second being whether area deprivation shapes neat risk. So these are the two main effects of adolescent mental health and adolescent area deprivation. And then the third research question is looking at this interaction between the two. So whether area deprivation modifies this relationship between mental health and mute risk. We were al also interested in whether these associations differ by gender. Which aspects of place might matter most? So if we look at the different dimensions of area deprivation, are there any that are particularly driving the effects? And finally, whether associations change across birth cohorts to see whether um the younger generations might be experiencing these worse than the older generations or vice versa. But for the purpose of today's talk, I'm only focusing on the first four research questions. Um we're yet to um complete research questions five and six. So the data that we used for this study was understanding society which was given a very nice introduction earlier in the day. Um it's a household longitudinal survey. um following 40,000 households from 2009 to 2025 and it collects um so much data and the specific measures that we've used for this study are for our exposure and psychological distress which we used the general health questionnaire which was um just nicely introduced by Kristoff um at age 16. Our moderator variable is area deprivation measured using the index for multiple deprivation which combines lots of different rankings from different domains of area level deprivation like income and education and health. Um ranks them and then um typically you present them in desiles with one being the most deprived and 10 being the least deprived. And we took this from age 16 as well because we wanted to try and get some of that adolescent um place context. And then our outcome is young people in the neat category between the ages of 18 to 24. We adjusted for um various confounders at age 16 as well. So that sort of baseline age that we've set some individual level factors like gender and ethnicity. some household level factors um like household income and then some parental factors including education, employment, social class and mental health because we really wanted to be able to sort of disentangle the higher level area level um socioeconomic factors from the individual level factors and we're also planning on um replicating the study using the millennium cohort study um which also got an introduction earlier in the day. So to be included in the analytical sample, participants needed to have observed >> Sorry, Katie. Sorry to interrupt. Your slides have just um stopped presenting. >> Oh, >> there we go. >> Oh. >> Oh, okay. >> Oh, >> is that Are they back? >> Yeah, they're back. >> Okay, cool. Um, yeah. So to be in the analytical sample and they needed to have observed psychological distress at age 16 and at least one neat observation between 18 to 24 and we also restricted this analysis to England because IMD isn't the most comparable across the different UK nations. Um to handle missing data we used multiple imputation by chain equations and the analysis that we used were mixed effect logistic regression models and then we calculated marginal effects afterwards and these models we used um to account for the panel structure of the data. So within the models we had just under 40,000 person age outcome observations from just under 6,000 um individual young people. And within those 40,000 observations, we had just under 7,000 observations um of being in the meat category, which was about 17.4% across ages 18 to 24. So our first research question was whether adolescent psychological distress shaped neat risk. So this is a line graph um of um yeah this relationship. So the horizontal um axis is baseline psychological distress score at age 16 with higher levels indicating higher levels of psychological distress and our vertical axis is the average predicted probability of being in the neat category. So what we can see is that as psychological distress increases so too does the average predicted probability of being in the neat category. So at the lowest levels of psychological distress, the predicted probability of being in any category is about 7% and this raises to about 13% at the highest levels of psychological distress. Our second research question was whether area deprivation shapes knee risk. So this is a similar graph but now our horizontal axis is baseline area deprivation um decile. So one is the most deprived area and 10 is the least deprived area. And what we can see here is that those in the most deprived areas um have an average predicted probability of being in the ne category of about 9% compared to just above 6% for those in the least deprived areas. So we're seeing this social gradient by area deprivation. Our third question was around whether area deprivation might modify the relationship between psychological stress and need. So this graph now the horizontal axis is back to being baseline psychological distress at age 16. Um but now we have two lines instead of one. The purple line is those it's representing the more deprived areas. Um and the blue line is representing the um less deprived areas. And the first thing to point out is that we can see that regardless of area deprivation, psychological distress is still associated with higher predicted probability of being in the neat category because both of these lines are going up as we move up the psychological distress scale. But what we can see is that these relationships differ by area deprivation. So at the lowest levels of psychological distress, we can see there's a clear distinction between the lines. So those in the mo in the more deprived areas have a predicted probability of being in the neat category of about 9% compared to 5% for those in the less deprived areas. Whereas when we move up that psychological distress scale we can see both of those um regardless of area deprivation are um falling around the 10%. Our fourth research question was around the gender differences in these associations. So this first graph um has it is looking at um how gender interacts with psychological distress to differentially influence needs. So um our horizontal axis is again psychological distress and our purple line this time is male and blue line is female. So what we can see is a very clear gender effect here. For females it doesn't really matter. It seems from this graph that um there isn't really as much of an association between psychological distress and predicted probability of being in the lead category. Um I think it's an increase of about 0.8% across the psychological distress scale. Whereas when we look at the male trend we can see that at the lowest levels of psychological distress this falls about six or 7% and this increases to 15% at the highest levels of psychological distress. When we look at the um gender differentiation for area deprivation, we can see that typically um the male trend is higher. So they're more likely to be um in the neat category, but these slopes are parallel between males and females. So um it doesn't really look like there's a gender interaction um for area deprivation. So just to quickly summarize the results for the main effects that we looked at, we found that young people with higher adolescent distress um and those living in more deprived areas at age 16 had higher predicted um meat risk. And then when we look at the um moderating effects, we saw that the psychological distress meat gradient appeared steeper in less deprived areas um and stronger among um men than women. It's important to say that with the um the area deprivation interaction um the effects were quite imprecise um because we it seems like we didn't really have enough power as we got to that higher end of the psychological distress scale. So thinking about the implications of this research um it suggests that early mental health support may have relevance for um need prevention. Um there's also uh it's really important to recognize that placebased disadvantage remains important and shouldn't be reduced to individual risk because we still saw um across the psychological dist distress scale those in the most disadvantaged areas had the highest um rates of being um in the meat category. And then also with these moderating effects that policy responses need to recognize different pathways into meat by gender and place. So maybe there needs to be more nuance around these um policies that are implemented. Quickly to go through the evaluation, it's got a longitudinal design which is really good for establishing temporality and reducing the risk of reverse causation. Um we also had repeated neat measurement which um is great because typically meat is measured at a single time point but we had it across 18 to 24 and we plan on disagregating those ages through marginal effects as well. We also use multiple amputation to try and keep as much power as we could and to account for some of the attrition bias um with loss to follow up. There is the possibility for residual confounding because our confounders were measured at the same time as the exposure um which meant um when preferably we would have them measured before. It also meant we couldn't account for some educational factors um which we would have wanted to because they would have been co-occurring at the same time as our mental health measure. um some of the interaction effects that the letter had limited power and a really important thing is understanding that neat is a really diverse category itself. So within neat you have those who are um unemployed and those who are economically inactive and the reasons for being in each of those is completely different and even within those is completely different. So we plan to um repeat these analyses disagregating that meat category further. So to conclude um we found that high adolescent psychological distress and area deprivation were associated with increased risk of being in the neat category with the um psychological distress meat association appearing stronger among men. Um and there are my references and thank you very much and I would love to answer any questions. >> Yeah, thank you very much Katie. That was that was fantastic. Um, we've already got a couple of questions straight in there. So, um, I had one I was going to ask as well, but I'll leave it to the audience first. So, we've got Annie asking, um, will this work feed into the government solutions work that you mentioned earlier? So, that's really interesting. Do you have a pathway um, to, you know, translate this research into some sort of policy impact? >> Yeah, that's a really good question. And one of the main things around equaliz the the center that I work for is around translating academic research into um solutions for health inequalities and we are currently working on a partnership with the um health foundation who are going to be um submitting evidence and we're planning on feeding these findings into their report um which is a collaboration with with us. Um so yes definitely we will be just thinking around um framing at the moment of the of the findings and hopefully um yeah it will be it will be used. >> That's great. Thank you. Another question from Ben um asking if you've controlled for educational attainment at age 16 and he'd assume that there's links to mental health and to you know being in the neat category. Um, and I think you're taking away some of the effect of mental distress, but useful to know how much goes independently into this. Um, thanks very much for a really interesting presentation. >> Yeah, that's a really good question and it's something that we had wrestled with for a while because um, we we didn't want to go into the youth survey for understanding society. We wanted to keep it as the adult survey. Um, so we haven't adjusted at the moment for anything education related at baseline. Um we that's why we want to do the um study or replicate the study in millennium cohort because we do have those earlier life measures that we can adjust for things like education. Um so no it's it's a really good point and we're definitely thinking about um considering that for for the next study. >> And I think I've got time just to slip in my question. I was going to ask about um you know area level deprivation is really interesting but I also wondered about um how often people move or whether they've moved areas. You know was that something you could look at? >> We haven't really um thought about that. It's a really good question. Um it would obviously add quite a lot of complexity to the analysis, but it's definitely something that should be looked at and it would be interesting to see whether someone moves from sort of a more deprived area to a less deprived area whether that takes any of the um effect away. Um yeah. No, we haven't we haven't really looked at that, but it would be interesting to >> Yeah. Well, that's great. Well, thank you very much, Katie. That was fantastic. Um next we have Sharon. So while Sharon's putting up her slides, I'll just introduce her. So Dr. Sharon Newfield is a welcome trust fellow and associate assistant research professor in the department of psychiatry at University of Cambridge. Her work seeks to clarify effectiveness of mental health services in young people and understand the course of depression and self harm over um adolescence. So her work has been cited in NHS funded in MRC strategy documents as evidence of good practice and also in clinical practice guidelines for treating depression. So maybe if you got time you can tell us how you did that Sharon because that's fantastic impact from your work. >> Okay. Thank you very much. Um and I'm really pleased to be able to speak about my team's work on understanding suicide attempt likelihood and um from depressive symptoms using the Millennium Cohort Study. And this is led by my PhD student CJ Lee who came uh to Cambridge as a visiting student from the University of Hong Kong. [sighs and gasps] So unfortunately suicide is the second leading cause of death among adolescents and it represents a important public health challenge around the globe. Um and by the time young people reach the age of 17 around 1 in 14 will have made a suicide attempt. Now, depression is one of the most robust predictors of suicide attempts in youths, but around twothirds of adolescence with depressive symptoms do not engage in suicidal behaviors. When you look at the UK NICE guidelines, they do not recommend screening for suicide risk, but they do recommend screening adolescence for depression. [snorts] and they the guidelines do underscore that when you're assessing depression risk, it's really important to consider a whole host of psychosocial factors. But among all those factors, it's really unclear which most strongly shape the link between depressive symptoms and adolescent suicide attempts. So what we want to do is try to um understand these pathways so that we can help improve prevention, early identification and targeted interventions. Now there has been some research which has tested the moderating pathways from depression to suicidality. Uh but this has really been limited to individual moderators. Um there's a lot of uh literature that supports that this that there's many more moderators that could contribute to this link, but nobody has been able to capture the complex interplay among multiple variables in this way. And this is largely because traditional statistical methods are quite limited. They're not able to detect complex multi-dimensional interactions or account for heterogeny or capture un underlying nonlinear relationships. And so some people have used machine learning to do this. And these algorithms can address those those um considerations that I mentioned on the previous slide. and they've been pred to to typically used in um testing how accurate a model is in predicting suicide risk. So basically it's a fancy regression model. You just see a really simple regression model here. So just imagine about 80 different X's instead of just the one that you see there and allowing for interactions and nonlinearities. That's what um machine learning models try to do. Um, and NICE guidelines have reviewed evidence from these machine learning models and they're saying they're still not good enough um to inform suicide risk assessment tools to to recommend suicide risk assessment tools because there's too wide confidence intervals around this sensitivity and specificity isn't good enough yet. Risk of bias. Um, so there are too many negatives to using this approach. But there is a newer approach um which is called double machine learning and this is actually based in the causal inference literature. So like other causal inference methods you can estimate an average treatment effect between a treatment or an exposure and an outcome. And in our data this is is then saying the the relationship between depression and suicide attempts. And this relationship is conditional on a whole host of observed coariants which you see as X on the figure there. Um and so then you can estimate the conditional treatment effects. And it's what's called a double debiasing method. Um and so just to outline this ever so briefly is that you have your um all your co characteristics, your predictors um are regressed onto depression and then separately in a model um they predict suicide attempts. And the residuals are obtained from these two models. And each of those models, you test out multiple different double machine learning methods to see which one operates best for the data that you have. And then in this stage two, you combine the residuals from the first two stages using the best model that was found. And you can do this to estimate the absolute risk in a population based on a treatment or an exposure. [snorts] So we were trying to use this method to understand how multiple risk and protective factors might contribute to suicide attempts in adolescence with depressive symptoms. Principally to help just identify what are those primary factors that link depressive symptoms to suicide attempts and h how do these this these factors interact to shape high and low risk of suicide attempts. So we used data from the UK Millennium Cohort Study as you have heard earlier today. This is a nationally representative birth cohort study. We focused on sweep six and seven age 14 and 17. And we used depresses depressive symptoms at age 17, suicide attempts at age 17 and risk and protective factors from age 14 and 17. Now um you might have noticed okay this is a contemporaneous model but we have longitudinal data. Why are you testing this model? Well the really what we have to you have to understand is that one of the assumptions of double machine learning is that there is not unmeasured confounding. And so because unfortunately with the millennium cohort study as great as it is there is this three-year time gap between measurements. And so we thought there would be too much risk of unmeasured confounding to acrew in that time that three-year age gap that we didn't think the models would work appropriately well. So we tested our primary model as one that is contemporaneous. But just to say is that um because of the the robustness of the double machine learning models, it is able to obtain causal inference even under um cross-sectional data. and at least we're able to use the covariants which are varying over time in this sample. So what are the risk and protective factors? So one thing some people might think about with machine learning is you just take a whole bunch of variables and you chuck them in. Uh it's very data driven and that can be a concern around it. Well for double machine learning model again because it's a causal inference method you have to be very careful about every variable that you choose to include in your model. there has to be a lot of literature that supports um its inclusion. So in the supplement of this paper um it there's this massive table that includes all the relationships between each one of these variables and depression and suicide attempts. Um so it they're strongly theoretically supported supported by previous literature to be included in this model as potentially having a moderating role between these uh depression and suicide attempts. So, we've got demographic features. We've got leisure activities and screen time. A lot of mental health variables, just one physical h health variable. We've got family and friendship, support. We've got substance use and gambling behaviors. We've got behavioral and psychological traits and and a whole suite of victimization experiences as well. So um this is quite robust in terms of covering psychosocial um possib uh moderate possible moderating factors and for the double machine learning that we've done in this paper we compared four different kinds of models um under different assumptions of linearity and sparsity with the causal force DML being the most nonlinear of all of the four models and each of these models consisted of training on 60% of the data and then validating that model on the remaining 40% and the models were evaluated with something called Ror which is a goodness of fit statistic like the R squared and this involves um five-fold cross validation to mitigate overfitting and that's like what you the kind of methods you would have heard in other machine learning um analyses. So then from these models we're able to estimate the average treatment effect and conditional effects and we interpret the conditional effects using the sharply additive explanations and tree interpreters which you'll see in a minute and we analyze this in Python. So Ca did this work my student um and uh we also thought it was really important to conduct sensitivity tests first of the average treatment effect of the main model. So we introduced a randomly generated confounder, unobserved confounders, and then re removed a random subset of the data. And all three of those sensitivity tests should have an average treatment effect that is very similar to the main one if the model is robust. And then finally, we tested a placebo effect to see whether that was close to zero. That was the hope. Um and then because our model is contemporaneous, we thought it was really important that we have a couple other sensitivity tests. We test for reverse causality. So whether suicide attempts might fit might predict depression at both at age 17 and whether 14year-old depressive symptoms could predict um suicidality at age 17. And this model could only include 43 coariantss because we could only include the age 14 coariantss compared to the 63 in the main model. So um also just a note on missingness in double machine learning it's important that this is used using complete cases as opposed to imputing because in this scenario imputing could introduce introduce estimation bias. um it could introduce data distortions and so we want to prevent these kinds of um errors and we also want to preserve the internal consistency of the data because the models rely on the sample sitting and uh sample splitting and crossfitting. So that then meant that it was important that we really think carefully about are we addressing the bias related to missingness. So um in the main analysis we're only able to use 70% of the sweep 7 data and so we thought let's see if we can minimize the uh missingness in another way. So all those 63 X variables the covariants that are the potential moderating variables we moved almost all of them into an auxiliary variable. So this means it's used for debiasing but it can't be used for the conditional model. So you can still estimate the average treatment effects. So we wanted to check whether the average treatment effects were the the same as the main model when we minimize the the missingness because in the the auxiliary variables missingness is modeled more in a fimmel kind of approach. So it's you're able to account for missingness more. Um so in this approach we're able to only have 15% missingness and we have really negligible effect sizes in all those 63 covariants um in terms of who's missing and not when we have our missing this sensitivity analyses whereas in the main analyses some of those effect sizes were small. So the hope is that the average treatment effects will be similar um across these two models. So when we do the model comparison, what we find is the best model overall captures nonlinear relationships. You can see what I've highlighted in green here is the causal forest double machine learning model that has the most appropriate ror model. All the other ones are negative and this actually suggests that there is overfitting in the model training stage. Um so it does not fit the data well. And when you look at the other sensitivity tests, um, we find that the average treatment effects are all basically the same, which supports the robustness of the model and the placebo treatment effect is close to zero, which is what we were hoping for. Um, we also find that the reduced missingness model has an average treatment effect that's almost identical to the main model, suggesting that the bias due to missingness is pretty small. Um and for our alternative models um we see that none of them are appropriate. All of them are negative. So this suggests that as we expected um the reverse model doesn't work based on theoretical um our understanding of of of how um depression and suicidality are related. um that supported um our our original um presupposition. And also our thought about how the data might be appropriately modeled was was also borne out in that they having this three-year time gap does not seem to um appropriately work in this data. And we suspect this is there's not enough coariants that we're able to include. There's not enough um confounding that we're able to adjust for. Um, so what does it mean to have an average treatment effect of 0.040? Well, what this means is that there's a one when there's a one standardized unit increase in the depression score. This means there's a 4% increased likelihood of a suicide attempt. Now, one standardized unit in our data represents 12 points on the K6, the depression measure that we're using, which has a clinical cutoff of 13 or more. So this means that for all but the lowest scoring participants when they transition from a non-depressed to a clinically depressed state there's going to be a 4% increased likelihood in a suicide attempt. Now you recall that I mentioned earlier that the average treatment effect rep represents an increase in absolute risk. So because again as I mentioned in the intro um there's a 7% prevalence in suicide attempts in 17year-olds in the UK this means that if we effectively intervene on depression in at this age range this could more than half the prevalence of suicide attempts in adolescence. So that just shows you how valuable even though it looks like a small number it's really has a lot of potential. So how do we model this heterogeneity? So this is the sharply added explanations chart um which on the the left panel panel A shows the average impact on the model input of the top 20 variables and then in the panel B on the right shows the distribution and the direction of each feature. So, you'll see, for example, um the long-standing illness, which is the first one, um if you have a long-standing illness, then it's going to be highly related to suicidality, but most people don't have a long-standing illness. Um so, and then how do we um visualize this with a three depth tree interpreter? You can see that on the left in figure two. We compare um and that has uh three levels of leaves um and is set to minimize overfitting with at least 100 group people in each leaf. Um and then we did the sub comparisons which you you can see in figure three. All but three of them are significantly different. And just to um focus in on the different profiles. So, the lowest risk profile is in purple. Um, and you can see that means that those with um high self-esteem and they hadn't experienced cyber victimization at age 17 and they had low depression scores at age 14, they had the lowest risk of suicidal attempts at age 17, um, which was only 2%. And then on the converse, the highest risk group was those who had low self-esteem, low mental well-being, uh, and a longstanding illness at age 17. And they had a 9% risk of suicide attempts. Um, but if you look at any of these um, groups at the very bottom, all of them are significantly different um, significantly higher than the lowest risk purple group. So if you have any one of these additional risk factors above and beyond depression, then that will um make you at higher risk for a suicide attempt. Um and so when you go back to the shops features, you can see that um those first five were the ones that we were looking at in in in the previous figure. Um but there's other ones that are important in the next five um important risk factors and protective factors that you can see there that I'll talk about and bring out in the discussion. Um so what does this mean for prevention? Um well first and foremost it really highlights the importance that we need to be investing as much as we can in early screening and intervening on depressive symptoms in adolescence. And we also need to be investing in those public programs that support families, that help families have as positive environments for children growing up because it really does confer um suicide resilience to suicide um in in young people. And if we can focus on establishing a safe digital environment for young people, this will have a big influence on um reducing suicide risk. And also um the educational programs that already exist. There's relationships and sex education which is um part of the department for education statutory guidance that schools should be following that already emphasizes relationships um respectful relationships, online risks, how to spot criminal behavior. So all this can really help in terms of ensuring that young people um know how to stay safe, know how to get help if something wrong happens, know how to communicate what's going on. So this we really need to be sure that we're we continue to focus on these programs um and because in fact all domains of victimization were in the top 20 of the of the most influential 63 coariantss um in our models. So what about implication for depression treatment itself? So because self-esteem was so important, we need to be sure that we we're including this in psychosocial assessments um and have this inform the treatment plans and there are already interventions for CBT that we need to really be ensuring that this is part of depression treatment. Um uh interventions that target uh self-esteem in particular, they also have good transdiagnostic impacts. So, not only is our study showing this is is likely to reduce suicidality, but it's been shown to improve well-being, decrease anxiety as well as depression. Um, and depression for treatment in early adolescence is really crucial. We saw that in that the 14-year-old data really predicted a higher risk of suicidality. um and we need to be sure that we're um targeting emotional symptoms and nurturing positive family um environments within the treatment environment. So those um all of those CAMS interventions that do involve um uh families in the in the treatment plan are really important to try to integrate as much as possible. We need to ensure that that trauma is part of our treatment planning um and processing those traumatic experiences because those were influential coariants and and as well we need to really be sure that we're integrating physical and mental health care in our the way that we design our services. So, um, you know, applauding the the the plans for for Cambridge, um, the new Cambridge Children's Hospital, which really is trying to think about how to integrate that really well, um, because we saw that long-standing illness was a really important factor. So, um, thank you everybody for your time and I really want to say thank you to CJ Lee uh, for her tremendous work on this um, and and, uh, Professor Paul Leip um, as well for his collaboration on this project. So, thank you everyone for your um time and attention as well. >> That's brilliant. Thank you very much, Sharon. I'm afraid um you go a little bit over those so we don't have much time questions. There is one in the question and answers box and if anyone else wants to put any quick questions in there um and there's one in the chat as well, but perhaps you can just reply to them in the question and answers and in the chat please Sharon um because I'd really want to move on to the next speaker. But yeah, really really interesting. So next we have um Tiara. So Tara, if you want to um share your slides while I introduce you. Um we've got um Tiara um Schwarz was Tofi who is a researcher who recently completed her infill and health data science at the University of Cambridge under the supervision of um Sharon who we just heard where she examined the association between cross uh reported mental health, symptom recognition and adolescence and service access. and she is now a neuroscience research assistant in the UK Dementia Research Institute. So that's fascinating. So that's great. So um I'll pass it over to you Tiara. Great. Thank you for the introduction. I'll go ahead and get started. So first to provide some background, the median age of onset of having any mental disorder is 14.5 years of age. Despite this, twothirds of children and young people with diagnosible mental health conditions do not seek help. And in the English context, of those who do seek help, less than a quarter seek specialist services, meaning services from professionals such as psychiatrists and psychologists who are specifically trained to treat mental health conditions. Now, why is that? Well, one huge barrier is a structural one, and that's the way that child and adolescent mental health services are designed in the United Kingdom. So essentially um CAMS is the service through which adolescence can seek mental health but uh but it's often through a referral process. So you often need to be referred most commonly by GPS but also by teachers or parents to seek mental health specialist services and access to self-referral. So referring yourself to mental health services within CAMS is not standardized across NHS trusts and many trusts which allow self-referral require parental consent under the age of 16. As a result, parents play a pivotal role in helpseeking among adolescence facing mental health difficulties. Specifically, 94% of young people attending an initial mental health intake reported that others influence helpseeking with parents playing the largest role and agreement on mental health symptoms between parents and children was associated with 43% of contact with a mental health provider for externalizing problems. In addition to parents, teachers are also important for service access. And this is for multiple reasons. So first, teachers are the most common source of help sought by parents on behalf of children. Uh and specifically, parental contact with teachers was associated with increased odds of contact with mental health specialist services on behalf of their child. In addition to the role that they play in supporting parents, teachers are key referers to CAMS. So even though GPS are the most common source of CAM's referrals, uh CAM's professionals reported that education professional referrals most often contain sufficient information for triaging and referral to specialist services more so than GP reports. So teachers are also very important refers to CAMS. Now we know that parents and teachers are important in helping adolescence gain access to mental health services. But what is unknown is whether the perceptions of adolescent mental health issues between parents, teachers and adolescence themselves um and whether these individuals agree. We don't know how that influences the probability of service contact. So that leads us to the central research question of this project, which is how do differences in how parents, teachers, and adolescence themselves perceive an adolesccent's mental health difficulties influence whether and from whom parents seek support for their adolescent. To examine this question, we use the mental health of children and young people survey, specifically the 2017 wave. And very briefly, it's a nationally representative probability sample drawn from the NHS patient register specifically to represent the English population. In terms of the measure of mental health that we used, we used the strengths and difficulties questionnaire. So in this survey, parents, adolescence, and teachers all filled out the strengths and difficulties questionnaire in reference to the adolescent emotional symptoms, conduct problems, peer problems, hyperactivity, and pro-social behavior using 25 items. to measure service contact in the age range we looked at which is 11 to 16 year olds early to mid-ado adolescence. We only had data available for parent reported service contact. So specifically parents were asked if they had been in contact with a range of different supports uh regarding the adolescent emotions, behavior or or difficulties in getting along with people. And we looked at four different levels of service contact. First informal contact constituting family and friends, the internet and telephone help lines. professional contact which refers to professionals that may be trained in youth mental health but not specifically trained in treating or specializing in mental health. So that refers to teachers, GPS and social workers. Mental health specialist contact which refers to psychiatrists and psychologists who are qualified to treat mental health issues and any contact which is an umbrella category encompassing helpseeking from any of these sources. Regarding our analytic sample, around 9,000 young people were included in MHCIP 2017. Of those, about a third were in the age range that we were interested in. And we decided to divide our sample into two analytic subsamples, specifically adolescence for whom the parent and the adolescent completed the SDQ and adolescence for whom parents and teachers completed the SDQ. And the reason we did this is because we're interested specifically in um in diatic analyses and to maximize the sample size as there was a large proportion of missingness in teacher SDQ reports in the sample. Regarding our methodology, we first used multi-group confirmatory factor analysis to determine whether items on the STQ measured the same construct across informants. So whether different informants interpreted questions similarly, which we found evidence for. and second to construct latent factor scores from a measurement model um in which we can reduce some of the measurement noise um involved in observed scores. Next, we use logistic polomial regression to examine the interactions between parent and teacher and parent and self-reported extracted latent SDQ subscores to predict parent reported service contact and we used estimated marginal means to evaluate the nature of any significant cross-informant interactions. In terms of missing data, we used complete case analysis since we were looking at some multi-informant data. Uh, regarding COVID adjustment, we adjusted for several sociodemographic factors as well as general parental distress since that's associated with both uh reports of adolescent mental health concerns and service contact. And for weighting, we created weights by multiplying the MHCP provided survey weights by inverse probability weights that we constructed for each analytic subsample based on the probability of complete SDQ data by informant. Now moving on to our key findings. So first we found that disagreement is common between informants on the SDQ. So to describe this figure here, we have a heat map basically crosstabulating uh parent SDQ scores against self on the left hand side and teacher on the right hand side SDQ scores across the three most clinically relevant SDQ subsklls. So that's conduct problems on top, emotional problems in the middle and hyperactivity at the bottom. So assuming that informants commonly agree, we would expect that this diagonal here would have the highest prevalence across the board. And while we do see that a majority um um we we do see that a large proportion of parent and self and parent and teacher pairs agree on the level of mental health concerns when um adolescence seem to be experiencing about average mental health concerns. Well, we don't see a clear clear diagonal pattern here. So there's quite a bit of disagreement as you can see by the prevalence on this side and this side of each of the different heat maps. So disagreement is common and it doesn't seem like there's a specific direction in which we see more disagreement um than the other way around. Second, there is a significant interaction between parent and self-reported emotional symptoms in predicting the odds of parent reported contact with informal and non-speist professional support. So to break down these interaction plots here on the x-axis we have parentr rated emotional symptoms of the adolescent on the y-axis we have the predicted probability of different of seeking different service contacts and we have three traces basically showing how parentrated emotional symptoms relate to service contact depending on self-rated emotional symptoms. So this orange trace shows um shows that relationship when the self when adolescence self-report below average symptoms. The teal trace shows when the adolescent reports average mental health symptoms and the purple trace shows when the adolescent reports above average emotional symptoms. And what we see here is that towards the first half of both of these graphs here, um, when the parent reports relatively average emotional symptoms in the adolescent, if the adolescent reports high emotional symptoms, as indicated by the purple trace, there's a higher probability that the parent will seek help for the adolescent compared to if the adolescent doesn't report mental health concerns or reports below average emotional symptoms. Now, this relationship seems to disappear as we get towards the higher end of parent rated emotional symptoms, suggesting that when parent rated emotional symptoms are high enough, it doesn't really matter what the adolescent self-reports, they're likely to seek informal contact or professional contact. Now, to quantify this, uh, we see that adolescence identifying emotional symptoms increases the odds of parents seeking informal or non-speist professional support if a parent has not identified emotional symptoms. So to break this down, when adolescence reported high emotional symptoms but parents did not, parents had around 52% higher odds of reporting informal contact, such as with family and friends, compared to pairs where neither reported emotional symptoms. Similarly, when adolescence reported high emotional symptoms and parents did not, they had 72% higher odds of reporting professional contact, such as with teachers or GPS, compared to pairs where neither reported emotional symptoms, suggesting that even if the parent doesn't identify issues in their adolescent. When it comes to informal and professional contact, the adolescent self-identifying emotional problems does increase the probability of service contact. Moving on to the parent teacher findings, what we see is that if a parent has already identified emotional symptoms in their child, parent teacher agreement greatly increases the odds of parents seeking any form of help. So these graphs look quite different to the parent self- relationships specifically in that teacher rated emotional symptoms seem to increase the probability of service contact across the board specifically general helpseeking. So this this relationship was significant for general helpseeking but not any of our specific contact types. And as you can see here with this purple trace where the teacher reports high emotional problems um this purple trace is higher than when the teacher does not report emotional symptoms or they report below average emotional symptoms for the adolescent. And this relationship seems to increase with increased parent reported emotional symptoms. So the the lines diverge more meaning that the probability of seeking help is greater is increased by a greater increment when the parent has already identified emotional symptoms in their adolescent. And to quantify this when both parents and teachers reported high emotional symptoms. So when there's agreement between parents and teachers, parents had twice the odds of reporting any service contact compared to parent teacher pairs where parents reported high emotional symptoms but teachers did not. Fifth, parents are largely the sole determinants of whether adolescence access mental health specialist services. So going back to these graphs, but focusing specifically on mental health specialist contact, you see that all of these traces overlap, suggesting that the adolescent reported emotional symptoms have no bearing on whether parents whether or not parents seek mental health specialist contact for their adolescent. It seems to be driven entirely by parentrated emotional symptoms. And we see the same relationship for teacher rated emotional symptoms. So it seems like parentrated emotional symptoms largely drive mental health specialist contact regardless [clears throat] of self-report or teacher report. Finally, you'll notice that we focused on emotional problems which is just one subscale of the STQ and that's because there were no statistically significant interactions observed for conduct or hyperactivity. Now moving on to the implications of these findings. So first what this underscores is the importance of parental mental health literacy. So across the board across subscales and between both of our samples parent reported concerns were the strongest predictor of parent reported service contact. So even though there are some interactions when it comes to seeking help in general, they're the strongest drivers. And for mental health specialist contact, they're the sole drivers of service contact at that level, which suggests that mental health literacy and parents is an essential area of intervention. And these efforts should facilitate not only parents identification of mental health concerns, but also their ability to evaluate the severity, significance, and the need for help to address mental health symptoms in adolescence. Second, our parent teacher findings underscore the importance of parent teacher communication. Specifically, teacher reports greatly increased general per parental helping seeking behavior and this relationship was stronger when the parent had already identified emotional problems in their adolescent suggesting that the corroboratory role of teachers is very important in encouraging service access. So, equipping teachers to effectively recognize and communicate emotional symptoms could encourage parental help seeking. But all of these implications assume that we're working within the current CAM system which largely require which largely depends on referrals by teachers, GPS or parents. What we could also consider is the importance of self-led pathways to mental health services. So considering parents are largely the drivers of access to mental health specialist services uh we may want to consider alternative models and specifically the WHO has expressed that services should offer multiple entry points and involve both youth and family participation and in different areas of the world especially in Australia they're trying out a youth single point of access model and I think recently they've expanded this to CAMS in the UK as well which proposes that there's a single point of access where youth can gain access to mental health services more easily without relying on referral or the tiered system. And emerging evidence suggests that these models are associated with improved mental health care and specifically improved mental health care access. So um this is another implication of our findings that potentially circumventing the issue of having teachers or parents mediate mental health access. It's possible that it'd be better to allow adolescence to seek these services directly. So that wraps up my presentation. Uh thank you so much for listening and uh I'd like to especially thank my supervisor Sharon for all of her help and support in the conceptualization of this project and I'll open it up for questions. That's brilliant. Thank you very much Tiara. Um uh we'll see if any questions come through in the question answer box. Just uh give it a minute to put some questions up there. Um but just while we wait for that, did you um you know did you manage to do anything with your findings like you know have you published them you know in a more sort of um open way you know to influence policy and practice or CAM support? >> Yeah, that's definitely on the horizon. So we're working to put this into a paper. Um there's some methodological knots that we're still untying, but um that's definitely the the hope that um this will provide some some vis that we'll be able to provide some visibility to this work, especially given the policy implications. >> Yeah, definitely. Brilliant. Um well, thanks very much. I don't think there's any other questions coming through right at the moment. Um but that's great. That means we can um move on on time to the next speaker. But if anyone has a question um for Tiara, feel free to put it in the question answers box and I'm sure she'll pick it up there. So, thank you very much. So, we'll move on to our final speaker for this session. Um um Verina um so while you're uh putting up your slides, I'll just read out your uh or introduce you. So um Verina um Schneider is a PhD student in the SockB center for doctoral training and biosocial research in the department of epidology and public health and a research assistant in the department of targeted intervention at UCL. Her research interests include social and health inequalities in the application of research and real world policy and practice and looking forward um to hearing your presentation. Um you just need to go full screen for us please Verina. [clears throat] Have I not done this? >> No. >> It's weird because it's coming up. Let me just um go again. H strange worked earlier. Let me just try this again. So maybe I just share the screen with the full presentation. Does that work? >> Yeah, perfect. >> Yeah. Sorry for all right. >> Okay. Um thank you very much for staying on to listen and what is very hot day today. Um my presentation is on sex and health across gendered context. A focus on generational differences. And um just yeah to start with um there are well-known sex differences in physical and mental health outcomes. And these health differences can be both due to biological sex such as chromosomal and hormonal factors as well as gendered social experiences. And the letter may include socioeconomic disadvantage, different exposure to psychosocial stressors and how they are embodied. And my PhD is um concerned with disentangling these different causes and mechanisms. And although that's quite challenging, it is quite important because we know that something is socially driven, we have an an opportunity to intervene. And this brings me to my terminology. So across this presentation when I refer to sex I refer to the biological distinctions um of chromosomal hormonal and anatomically female or male characteristics and acknowledge these are not strictly binary um but these are um often how they are the data are recorded in available research data but I'm also focusing on this external binary world of binary gender norms and therefore um gendered life course patterns that are associated ated with a person's perceived or assigned sex at birth. And accordingly when I talk about gendered mechanisms and pathways and um gender I'm talking about this also in this binary downstream lift experience. So there are some conceptual and methodological complexities when it comes to this entangling sex and gender. And this is because gender isn't really well defined or measured. It's not one variable. um and it's quite bearable both within person and also between contexts and it's quite multi-dimensional. Um if we look at the causal inference framework that poses some challenges to some assumptions where we don't really have a consistent effect of the exposure on the outcome um because gender is not clearly defined and there's not really a manipulable treatment in a strict sense. And then um in terms of positivity a challenge is that one cannot really experience the gendered environment associated with another sex assignment at birth. So in my PhD I'm using a trioulation of different methods to approach what is social and what are the gender mechanisms in the sex and health association. And my talk today will look at a heterogenity analysis. And what this is um I'm using heterogenity analysis to compare a causal the causal effect of sex on health across more or less gendered contexts. And here I'm using generation as a proxy for social gendered context because we know that gender norms and gender attitudes have become uh progressively more gender equal over time. That has been attributed to generational replacement rather than age or period effects alone. Um so basically um what I'm saying here is if I don't assume biological differences between men and women have changed over generations. So if we do say see the sex health association change over generations we may assume something social is going on. Just looking at the literature and this is a little wider the literature um coming from a different perspective but um the consistent finding when we're looking at um outcomes smoking, alcohol consumption and lung cancer mobility or mortality which are outcomes with male disadvantage um the sex differences are narrowing in younger generation and this has been attributed to um health behaviors in women peaking later and therefore the lung cancer risk um peaking in later cohorts um which is um are sit with the smoking um incline and then for heavy life expectancy mobility and multimobidity um which is a female disadvantage um we see increasing sex differences um across two studies for younger cohorts for mental health the um evidence is quite consistent uh we see studies that show both narrowing and wid widening disparities in depression. Um widening sex gap in suicide mortality um and other studies that show um their psychological sex differences have been quite stable over time and also for self-rated health the evidence is very inconsistent. Um so overall this inconsistent evidence um may be due to varying country contexts um sample characteristic but most importantly there are quite a lot of variations on how studies have dealt with an issue I'm going to come into uh more detail uh which is the age period cohort issue if we're interested in cohort or generation effects we need to also account for age and period and there are some issues that have um not been addressed explicitly across all studies or assumption haven't been made clear and um no study was identified in the UK context um on this matter. So my aim was to examine sex differences in mental health, physical functioning and poor self-rated health and the effect modification by generation within a large longitudinal study representative of the UK population. And first I'm asking how is sex above associated with these health outcomes? And second um is the association modified um by different gendered um contexts such as generation. Um I'm using understanding society data um which has been introduced before but quickly um it's a um nationally representative household survey um of over 40,000 households and I'm using the harmonized data set which includes the BHPS which the former British household panel survey so of data starting from 1991 till 2024 when I conducted the study um which is quite good Um for um the age period cohort um question here my exposure is sex at birth but it's important to acknowledge this is not administrative data. This is self-reported by a household member. And then my effect modifier is generation which is categorized in three um categories by the year of birth. Um so I'm having mainly baby boomers in the first categories and a few of silent generation. The second category is um generation X and the last category generation Y and Z anyone born after 1980. My um outcome measures for mental health are the GHQ12 which I focus in this presentation. Um this is has been measured across all ways of the BHPS and UK hls but I also analyze the SF12 mental component summary. Um physical functioning is measured also with the SF12 physical component summary. And finally one item of the SF12 um is used for binary um measure of poor self-rated health. included were men and women between the ages of 16 and 65 years old who had a non-m missing outcome at baseline and at least two follow-ups and inconsistent um reports of sex were excluded due to the definition of sex. Here um I've used group of models in STA with cubic age trajectories, random intercepts and random slopes for age and clustering at PSU level. I've used the provided um baseline survey weights and to interpret the effect sizes I use marginal effects to estimate the average effects of sex. So when it comes to the moderation analysis I alluded to already h there is the age period cohort short APC identification problem. What this means is that whenever we have periods in my case that's wave um chord in my case generation and age these are linear dependent and although I'm not interested in the main effect of cord or generation this also applies to the um interaction effects or the product term and um there's not really a solution unless we learn to time travel um but in order to run models we can make some untestable assumption and commonly and often not explicitly um done is to emit omit one of the factors from the model. Um so we don't run into the multiolinearity issue anymore. Um but that um implicitly implies that if we let's say we drop period from the model, we say that the period effect is zero. And if that's not true, then the model suffers from emitted variable wires. A little bit of a weaker assumption can be done in three factor models. And really what this does is um it only the identification problem only applies to linear effects. If we do nonlinear transformation to our variable um we don't have the problem any longer. So that means we do we can model nonlinear effects of let's say period and only excludes the linear effects. Um, so that's a weaker assumption that only applies to the linear effects. And I in my analysis I've done both just to see the effects um of the bias, but I'm um concentrating on the three factor models. Again, I estimate the marginal effects um of sex. And here I'm using ages that are observed across all three generations. And I'm showing in a moment why that is and why that is important. So here we see the sample sizes. Um I had over 56,000 for GHK12 and over 44,000 for the SF12 scales. Um and quite a nice split across the generations and um sex. [snorts] And here I'm showing the age distribution by generation to um show what I mean and what what needs to be acknowledged um in the analysis. So we observing the youngest generation at younger ages and the oldest generation at older ages and if we just predict our outcome using those distributions we may just see the effect of age and not that of generation. Um so therefore in my um study when I predict outcomes I'm using ages 26 to 42 um for example for GHQ 12 um to make sure I um only predict ages that were actually observed. So my first question was how is sex at birth associated with health outcomes? So across ages and waves, women experience higher distress across um all time points. And this graph shows the predicted GHQ12 score over um age um for women and men. And um GHQ2 is measured on a scale from 0 to 36 with higher scores meaning higher distress. So we can see women on average score 1.26 higher. Women also have a higher probability to report poor self-rated health across all waves and ages and that's um 2.4% higher probability. Um there wasn't really a meaningful gender gap in physical functioning. So this is a Z standard uh standardized score. So we're talking about 0.08 standard deviations um here. So second then I'm looking at how this relationship varies by um or cross generation and first of all I just want to show what happens if we don't account for age. So we can see here the baby boomers are observed later than um generation X and generation Y and Z. Um so what we see here the sex difference in GHQ12 uh increases from 1.08 08 to 1.14 and 1.60 as we move to the younger generations. But what we don't really know how much of this is driven by greater sex differences in younger ages which we can see here in the growth curve models. So once we run or run the three factor models and restricted the ages we do see an attenuation of the effect but interestingly we see still the see the same trend that we see an increase in the sex differences in younger generation from 0 1.0 0 to 1.3 in the youngest generation. For um self-rated health um I had to now do pairwise comparison because that's only measured in the UK hls. So I had less of an overlap. So I'm first comparing baby boomers in generation X at the ages of 45 to 58. And I see an increase in sex difference of 1.9% to 4.3%. And then if you look at the younger generations um at ages 29 to 42 I'm seeing an increase of 1.3% sex difference in generation X to 3.0 in generation Y and Zed. And then finally although the overall differences in physical functioning weren't meaningful we do see the same story here. So we see an increase um in the younger generations across both age comparisons. So in summary, women consistently report poorer mental health and self-rated health than men and these differences are more pronounced in the younger generation groups despite more egalitarian gender norms. Um however they are small overall. Um there was no meaning sex difference in physical functioning. However, with the differences we saw across generations followed the same trend as in the other outcomes and I haven't shown sensitivity an analysis but the results were quite robust. So we can't really interpretate the magnitude of the generation differences as an effect of um gender and that's because we don't really observe a generation that is completely gendered versus one that isn't um where gender processes do not exist. As I said at the start gender is also quite multi-dimensional. So there may be opposing social processes. So this study hasn't really looked at the actual mechanisms. We're just looking at um a uh yeah a way to approach it um conceptually. Um and although we see only small differences across generation, this may be interpreted as evidence for gendered and social mechanisms. There are a few limitations I just quickly want to go through. I mentioned the ABC issue and um the potential for bias. my assumption that there are no linear period trends may be wrong uh and there may still be a bias. I also used wave to approximate um period but um the way the data are collected there um there is some noise in there it's not exactly mapped to year um and then we have a limited age overlap for some of the outcomes and the choice of generation it's a categorization so every categorization is a bit arbitrary so that's um also limitation as I said the study doesn't really know it's a mechanism so that will be next steps to look into that I talked about the magnitude of the effects that we can't really interpret. Um I also haven't um the study hasn't looked at within gender inequalities or intersectionalities and um I've only used baseline survey weights um because the longitudinal provided weights would have cut my sample by quite a bit. So that there is some potential for attrition bias. However, we've done some sensitivity analysis with a smaller sample and the longitudinal weights and the results were the same. And finally, we're talking about self-reported outcome measures and we need to acknowledge that there are quite generate reporting um of health outcomes as well. So, this brings me to the end of my presentation um and I'm looking forward to your questions.