Trajectories of change after a health-education program in Japan: decay of impact in anxiety, depression, and patient-physician communication

Background Health education can benefit people with chronic diseases. However, in previous research those benefits were small, and reinforcement to maintain them was not effective. A possible explanation is that the benefits appeared to be small and reinforcement appeared to be ineffective because those analyses mixed data from two latent groups: one group of people who needed reinforcement and one group of people who did not. The hypothesis is that mixing the data from those two different groups caused the true effects to be “diluted.” Methods To test that hypothesis we used data from the Chronic Disease Self-Management Program in Japan, focusing on anxiety, depression, and patient-physician communication. To identify latent trajectories of change after the program, we used growth-mixture modeling. Then, to find out which baseline factors were associated with trajectory-group membership, we used logistic regression. Results Growth-mixture modeling revealed two trajectories—two groups that were defined by distinct patterns of change after the program. One of those patterns was improvement followed by backsliding: decay of impact. On anxiety and depression the decay of impact was large enough to be clinically important, and its prevalence was as high as 50%. Next, logistic regression analysis revealed that being in the decay-of-impact group could be predicted from multimorbidity, low self-efficacy, and high scores on anxiety or depression at baseline. In addition, one unexpected finding was an association between multimorbidity and better patient-physician communication. Conclusions These results support the hypothesis that previous findings (i.e., apparently small effect sizes and apparently ineffective reinforcement) actually reflect “dilution” of large effects, which was caused by mixing of data from distinct groups. Specifically, there was one group with decay of impact and one without. Thus, evaluations of health education should include analyses of trajectory-defined groups. These results show how the group of people who are most likely to need reinforcement can be identified even before the educational program begins. Extra attention and reinforcement can then be tailored. They can be focused specifically to benefit the people with the greatest need.


Study design and timing of measurements 124
Data were collected four times over one year. Baseline data were collected before the 125 first group-discussion session. Follow-up questionnaires were sent by postal mail 3, 6, and 12 126 months later. A post-paid self-addressed envelope was included. A reminder postcard was sent to 127 each participant whose follow-up questionnaire was not received within two weeks. 128

Analyses 129
To allow detection of decay of impact, the analyses were done using data from 130 participants who provided at least three waves of data (456/643; 71%). Unconditional quadratic 131 Growth-curve analysis indicated that for all three outcomes the quadratic terms were 163 significant (see Table 2). In addition, higher self-efficacy at baseline was associated with less 164 anxiety, less depression, and better communication with physicians. 165 Also associated with both anxiety and communication was the number of diagnoses, and 166 the regression coefficients for both were positive. That is, the participants with more comorbid 167 conditions had greater anxiety at baseline. It is noteworthy that the participants with more 168 comorbid conditions also had better baseline scores on the scale measuring communication with 169 physicians. These associations did not change over time, as evidenced by the non-significant 170 interaction term with time. 171

Groups defined by their trajectories 172
For all three outcomes, the GMM results were similar: The BIC and the P c both led to the 173 conclusion that the best-fitting models were those with two groups (Appendix 2). 174 For each outcome, those two groups began from substantially different baseline scores 175 ( Figure 1). Also for each outcome, one group changed very little throughout the follow-up period 176 while the other changed noticeably within the first six months of follow-up, and then it reversed 177 course back toward the baseline value ( Figure 1). That is, on each outcome some participants 178 were in a decay-of-impact group and the others were not. About half of the participants were in 179 the decay-of-impact group: anxiety 45.6%, depression 50.7%, communication with physicians 180

46.3%. 181
Participants who had decay of impact on one of the two mental-health outcomes (anxiety 182 or depression) were also likely to be classified as having decay of impact on the other one (Phi = health outcomes were no more or less likely to have decay of impact on communication (Phi = 185 0.095 and 0.043, Appendix 3). 186 On the two mental-health outcomes, the decay-of-impact group was the group with worse 187 baseline status: more symptoms, and more-frequent symptoms, of anxiety and depression. In 188 contrast, on communication with physicians the decay-of-impact group was the group with better 189 baseline status: more frequent use of the three specified methods for good patient-physician 190 communication. Also, by the end of the follow-up year the anxiety and depression scores had 191 decayed back to their respective baseline levels, whereas on communication the decay trajectory 192 was clear but the scores did not return to the baseline level, in other words the decay itself was 193 smaller on communication with physicians than on the mental-health outcomes ( Figure 1). 194

Contributors to group membership (Table 3) 195
For all three outcomes, self-efficacy at baseline was associated with group membership. 196 Participants with higher self-efficacy were more likely to be in the group with lower anxiety at 197 baseline, in the group with lower depression at baseline, and in the group with better 198 communication at baseline. 199 For anxiety and for communication with physicians, the number of diagnoses was also 200 associated with group membership. Regarding anxiety, participants with more diagnoses were 201 more likely to be in the group with higher (i.e. worse) scores at baseline and subsequent decay of 202 impact. Regarding communication, participants with more diagnoses were more likely to be in 203 the group with higher (i.e. better) scores at baseline and subsequent decay of impact. 204

All participants 206
When all participants were considered together, all three outcomes improved over the 207 first six months. That improvement was followed by a small deterioration. Thus, even from the 208 least-detailed analyses, some decay of impact was evident (Table 1). That interpretation is 209 supported by the results of the growth-curve analyses: the quadratic terms were significant for all 210 three outcomes (Table 2). 211 The growth-curve analyses also showed that higher self-efficacy at baseline was 212 associated with less anxiety, less depression, and better communication with physicians, which is 213 consistent with the theoretical basis of the CDSMP (Lorig & González, 1992). 214 Also evident at this level of analysis were associations with multimorbidity. Having more 215 diagnoses was associated with more anxiety, more depression, and better communication with 216 physicians. Of those three findings, the first two might be expected, but the third is particularly 217 interesting. It is also reflected in the group-membership analyses, and so we will discuss it 218 below. 219

Findings from GMM: groups defined by their trajectories 220
For all three outcomes the results of GMM were consistent: Among all of the models 221 tested, the two-group models had the lowest BIC and the highest P c . We are therefore confident 222 in saying that GMM revealed two latent groups among these participants in the CDSMP. In some 223 circumstances practical considerations could override the conclusions from those statistical 224 criteria, as described in Appendix 2. 225 As noted above in the Introduction, in some previous studies, subsets of CDSMP 226 participants were defined a priori and with reference to theory. In contrast, groups identified by GMM are empirical, as is each participant's group membership. We note that the GMM 228 approach can lead to testable hypotheses (regarding multimorbidity, as described below), and it 229 can be used to answer important questions about whether similar phenomena also occur among 230 other groups and in different settings. 231

Mental health, and reinforcement 232
For both mental-health outcomes, the trajectory-defined groups differed in their baseline 233 status and in their pattern of change after the program. Regarding anxiety, approximately half of 234 the participants were in a group that began with relatively good scores, and they improved very 235 gradually over the following year. In contrast, the other half were in a group that began from a 236 high-anxiety baseline. That second group improved over the first 6 months, but by the time of the 237 12-month follow-up it had returned to its baseline level, and thus we refer to the latter group as 238 the decay-of-impact group. The same was true with regard to depression. 239 Dichotomization is undoubtedly dangerous (Harrell, 2019), and yet HADS scores are 240 used to separate people into categories of anxiety and depression severity. In Japan, the HADS 241 threshold score separating non-cases from possible and probable cases was 9 (Matsudaira et al., 242 2009). The decay-of-impact trajectories on both anxiety and depression crossed that threshold 243 twice -first during the improvement occurring soon after the intervention ended, and then again 244 about 8 months later during the decay back toward the baseline value (Figures 1a and 1b). 245 Therefore, to the extent that the threshold of 9 is useful, both the improvement measured soon 246 after the CDSMP and also the deterioration measured near the end of the follow-up year were 247 clinically important. group and the movement of that group between clinical categories support the idea that follow-252 up interventions -reinforcement -should be offered to some of the participants. Had 253 reinforcement been given to all, it is unlikely that those in the group without decay of impact 254 would have benefitted from it, simply because they already had almost no psychological distress 255 -almost no room to improve. Rather than being expended on all of the participants, the 256 resources used to implement reinforcement should be saved for the people who need it, to help 257 them maintain their newly-improved status or perhaps improve further. 258 The present findings show how the small effect sizes and null results in published studies 259 of reinforcement could be underestimates. Specifically, if all participants are considered together 260 then any benefits to the decay-of-impact group will be diluted. Testing the effects of 261 reinforcement is reasonable, but only in those whom reinforcement could benefit -the decay-of-262 impact group. To identify that group at baseline, i.e., even before the CDSMP begins, the present 263 findings offer two potential criteria: a high baseline score (greater distress) on the mental-health 264 outcome of interest, and a low level of self-efficacy. With regard to anxiety, a third criterion 265 could be multimorbidity. Also noteworthy are the three criteria that might be used to pre-emptively identify the 279 participants who are most likely to need communication-skill practice: a low communication 280 score at baseline, low self-efficacy, and unimorbidity. 281

Multimorbidity 282
The participants with more diagnoses had better communication scores in the initial 283 growth-curve analysis (Table 2), and they were more likely to be in the trajectory-defined group 284 that had better communication scores throughout the year (Table 3). While the CDSMP has been 285 found to be particularly useful to people with multiple diagnoses (Harrison et al., 2012), here 286 multimorbidity was associated with a desirable health-related behavior even at baseline. To 287 address the apparent connection between having multiple diagnoses and communicating well 288 with physicians, we begin by noting that the communication scores reflect how often the 289 respondents do the following three activities: making a list of questions to ask one's physician 290 during clinic visits; asking one's physician about things that one wants to know or does not 291 understand regarding one's treatment; and discussing (with one's physician) personal problems 292 related to one's medical condition. The people with multiple diagnoses probably had more 293 experience being in health-related situations that were difficult to manage. To deal with those 294 difficulties, perhaps they began writing lists of questions, asking for clarification, and discussing 295 personal issues related to their diseases, simply because their health conditions were so complex.
We hypothesize that at least some of the people with multimorbidity had become accustomed to 297 doing those activities, and therefore in the domain of patient-physician communication they had 298 already become "expert patients" (Reeves et al., 2008) by the time the study began. To the extent 299 that better patient-physician communication results in better clinical care, this connection 300 between multimorbidity and good communication could account at least in part for the 301 documented association of multimorbidity with higher-quality care (Min et al., 2007). 302 Other factors could also be important. For example, self-selection might have played a 303 role. After all, participation in the program was voluntary (as it is worldwide). Personality can 304 moderate the effects of the CDSMP (Franks et al., 2009;Jerant et al., 2010), and perhaps it also 305 affects one's decision to participate. Among all of the eligible people with multimorbidity, those 306 who are less "conscientious" and less interested in self-managing their conditions would not 307 often make lists of questions, etc., and they might not have found the CDSMP to be attractive 308 and thus would not have participated. In contrast, the CDSMP might appeal to people like the 309 highly-proactive communicator with 8 chronic conditions who was described by Haslam (2015). 310 People with multiple diagnoses who take initiative in self-managing their condition(s) by writing 311 lists of questions, etc. could be over-represented among the program's participants. 312 As explanations of the association between multimorbidity and good patient-physician 313 communication, both the self-selection hypothesis and the already-an-expert-patient hypothesis 314 remain to be tested, and of course both could be true. 315

Limitations 316
The four waves of data collection over one year were more than enough to allow 317 detection of decay of impact, but more frequent measurement and longer follow-up would of 318 course be useful. 319 TRAJECTORIES AFTER HEALTH EDUCATION: DECAY OF IMPACT 16 The number of diagnoses was self-reported. While we would have preferred to use 320 medical records, for many chronic conditions self-reported diagnosis is accurate enough for 321 research (Karison et al., 1999;Wada et al., 2009). 322 Summary 323 GMM exposed two trajectory-defined groups, and the CDSMP clearly benefitted one 324 group more than the other. However, the group that benefitted also had substantial decay of 325 impact, and thus needed reinforcement. The decay-of-impact group comprised almost half of the 326 participants. At baseline (i.e., before the program began), the participants most likely to need 327 reinforcement were those with multimorbidity, those with low self-efficacy, and those who were 328 clinically anxious or depressed. Smeulders, E., van Haastregt, J., Ambergen, T., Stoffers, H., Janssen-Boyne, J., & Uszko-471 Lencer, N., Gorgels, A., Lodewijks-van der Bolt, C., van Eijk, J., & Kempen, G. (2010). 472 Heart failure patients with a lower educational level and better cognitive status benefit 473 most from a self-management group programme. Patient Education And 474 Counseling, 81(2), 214-221. doi: 10.1016/j.pec.2010.01.003 475 backgrounds. Patient Education And Counseling, 64(1-3), 360-368. doi: 479 10.1016/j.pec.2006.04.003 480 Wada, K., Yatsuya, H., Ouyang, P., Otsuka, R., Mitsuhashi, H., & Takefuji, S., Matsushita, K., 481 Sugiura, K., Hotta, Y., Toyoshima, H., & Tamakoshi, K. (2009. Self-reported medical 482 history was generally accurate among Japanese workplace population. Journal Of 483 Clinical Epidemiology, 62 (3), 306-313. doi: 10.1016/j.jclinepi.2008.04.006 484 Whitelaw, N., Lorig, K., Smith, M., & Ory, M. (2013. Webinar: Findings from the CDSMP 485 National Study. Retrieved from https://www.ncoa.org/resources/webinar-findings-from-486 the-cdsmp-national-study/ 487 Figure 1. Trajectories of change after health education, showing two trajectory-defined groups for each of the three outcomes Growth-Mixture Modeling revealed two trajectory-defined groups for each outcome. On anxiety and depression higher scores are worse. On communication with physicians higher scores are better. For each outcome, one of those two groups had improvement followed by deterioration: decay of impact. For Anxiety and Depression, a score of 9 is the cutoff used in Japan to separate non-cases from possible and probable cases. The outcomes discussed here are anxiety, depression, and communication with physicians. Self-efficacy was used as a mediator in subsequent analyses because of its importance in the theoretical basis of the CDSMP.  The results shown are from logistic regression. The 0-1 coding of group membership (which is the dependent variable) reflects the relative magnitudes of the baseline scores. For all outcomes, the group with the lower baseline score is coded "0" and the group with the higher baseline score is coded "1." Thus, the group with less anxiety at baseline is coded "0" while the group with more anxiety at baseline is coded "1." The same is true for depression. In contrast, the group with better communication (higher scores) at baseline is coded "1" and the group with worse communication (lower scores) at baseline is coded "0." Appendix Figure 1c. Frequency distribution of number of diagnoses Appendix Y = 4.3141 + 0.3974*X -0.0322*X 2 Y = 5.9052 -0.4869*X + 0.0282*X 2 Y = 6.9689 + 0.3035*X -0.0105*X 2 Y = 11.8108 -1.0304*X + 0.0731*X 2 4 Appendix Y = 2.9075 + 0.1054*X -0.0087*X 2 Y = 6.2069 -0.0338*X -0.0034*X 2 Y = 7.065 + 0.6366*X -0.0289*X 2 Y = 8.7784 + 0.1061*X -0.0103*X 2